repo stringclasses 1
value | github_id int64 1.27B 4.87B | github_node_id stringlengths 18 24 | number int64 8 14.2k | html_url stringlengths 49 53 | api_url stringlengths 59 63 | title stringlengths 1 402 | body stringlengths 1 62.9k ⌀ | state stringclasses 2
values | state_reason stringclasses 4
values | locked bool 2
classes | comments_count int64 0 99 | labels listlengths 0 5 | assignees listlengths 0 5 | created_at stringdate 2022-06-09 16:28:35 2026-07-12 15:40:16 | updated_at stringdate 2022-06-12 22:18:01 2026-07-12 20:19:03 | closed_at stringdate 2022-06-12 22:18:01 2026-07-09 13:56:15 ⌀ | author_association stringclasses 3
values | milestone_title stringclasses 0
values | snapshot_id stringclasses 132
values | extracted_at stringdate 2026-04-07 13:34:13 2026-07-12 20:35:20 | author_login stringlengths 3 28 | author_id int64 1.54k 296M | author_node_id stringlengths 12 20 | author_type stringclasses 3
values | author_site_admin bool 1
class |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
huggingface/diffusers | 4,613,412,160 | I_kwDOHa8MBc8AAAABEvsVQA | 13,887 | https://github.com/huggingface/diffusers/issues/13887 | https://api.github.com/repos/huggingface/diffusers/issues/13887 | When use UniPCMultistepScheduler as the ODE solver, and input different num_inference_steps in StableDiffusionXLPipeline, the first inference step is wrong | ### Describe the bug
When the user calls the StableDiffusionXLPipeline using the ckpt from stabilityai/stable-diffusion-xl-base-1.0 and sets UniPCMultistepScheduler as the inference schedule, if the user only inputs num_inference_steps without setting timesteps, the inference time schedule would be incorrect. Setting... | open | null | false | 1 | [
"bug",
"needs-code-example",
"pipelines",
"schedulers"
] | [] | 2026-06-08T13:44:32Z | 2026-06-14T23:17:54Z | null | NONE | null | 20260614T233521Z | 2026-06-14T23:35:21Z | TheLovesOfLadyPurple | 227,429,101 | U_kgDODY5K7Q | User | false |
huggingface/diffusers | 4,626,368,831 | I_kwDOHa8MBc8AAAABE8DJPw | 13,903 | https://github.com/huggingface/diffusers/issues/13903 | https://api.github.com/repos/huggingface/diffusers/issues/13903 | [Anima] Add img2img capability | **Is your feature request related to a problem? Please describe.**
Anima in diffusers is a modular diffusers only model but is still missing the image to image functionality. This should be relatively easy to add and can help people understand how to work with modular diffusers.
This feature request is to see if some... | open | null | false | 4 | [
"good first issue",
"contributions-welcome",
"modular-pipelines",
"feature-request"
] | [
"PreethamNoelP"
] | 2026-06-09T23:11:19Z | 2026-07-07T17:09:37Z | null | MEMBER | null | 20260707T173513Z | 2026-07-07T17:35:13Z | asomoza | 5,442,875 | MDQ6VXNlcjU0NDI4NzU= | User | false |
huggingface/diffusers | 4,630,193,172 | I_kwDOHa8MBc8AAAABE_skFA | 13,907 | https://github.com/huggingface/diffusers/issues/13907 | https://api.github.com/repos/huggingface/diffusers/issues/13907 | Support for JoyAI-Echo | ### Model/Pipeline/Scheduler description
# Model/Pipeline/Scheduler description
**JoyAI-Echo** is a framework designed to address key challenges in long video generation, including error accumulation, weak temporal coherence, and high latency, which limit its usability in interactive scenarios. It tackles these issues... | closed | completed | false | 1 | [
"pipelines",
"feature-request"
] | [] | 2026-06-10T10:14:27Z | 2026-06-10T12:07:08Z | 2026-06-10T12:07:08Z | CONTRIBUTOR | null | 20260610T123520Z | 2026-06-10T12:35:20Z | feice-huang | 105,336,512 | U_kgDOBkdOwA | User | false |
huggingface/diffusers | 4,631,081,242 | I_kwDOHa8MBc8AAAABFAixGg | 13,909 | https://github.com/huggingface/diffusers/issues/13909 | https://api.github.com/repos/huggingface/diffusers/issues/13909 | [New Pipeline/Model] Add JoyAI-Echo multi-shot audio-video generation pipeline | ### Model/Pipeline/Scheduler description
**We are the JoyAI Team (JD.com)**, proposing the integration of JoyAI-Echo into Diffusers.
**JoyAI-Echo** is a unified framework for long-form audio-visual generation, designed to support minute-level video creation with synchronized audio, strong temporal consistency, and re... | open | null | false | 0 | [
"models",
"pipelines",
"feature-request"
] | [] | 2026-06-10T12:14:31Z | 2026-06-10T12:15:23Z | null | NONE | null | 20260610T123520Z | 2026-06-10T12:35:20Z | sjq66 | 142,070,647 | U_kgDOCHfTdw | User | false |
huggingface/diffusers | 4,639,181,759 | I_kwDOHa8MBc8AAAABFIRLvw | 13,920 | https://github.com/huggingface/diffusers/issues/13920 | https://api.github.com/repos/huggingface/diffusers/issues/13920 | Ideogram4: `Ideogram4MRoPE` breaks under `torch.autocast`: all image positions collapse, producing flat single-color images | ### Describe the bug
`Ideogram4MRoPE.forward` computes the rotary frequencies with a matmul:
```python
freqs = inv_freq @ pos.unsqueeze(2)
```
`torch.matmul` is on autocast's lower-precision op list, so when the transformer's forward runs inside
`torch.autocast("cuda", torch.bfloat16)` — common in training code, an... | closed | completed | false | 2 | [
"bug",
"models"
] | [] | 2026-06-11T09:36:36Z | 2026-06-19T22:02:32Z | 2026-06-19T22:02:32Z | CONTRIBUTOR | null | 20260619T223512Z | 2026-06-19T22:35:12Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,645,383,278 | I_kwDOHa8MBc8AAAABFOLsbg | 13,925 | https://github.com/huggingface/diffusers/issues/13925 | https://api.github.com/repos/huggingface/diffusers/issues/13925 | 1 bit, ternary image generation support | **Is your feature request related to a problem? Please describe.**
Currently no support for 1 bit image generation
**Describe the solution you'd like.**
[1-bit and Ternary Bonsai Image 4B: Image Generation ](https://prismml.com/news/bonsai-image-4b)
**Describe alternatives you've considered.**
https://github.com/Pris... | open | null | false | 1 | [
"quantization",
"pipelines",
"feature-request"
] | [] | 2026-06-12T01:12:07Z | 2026-06-12T17:57:39Z | null | NONE | null | 20260612T183524Z | 2026-06-12T18:35:24Z | rupeshs | 3,255,994 | MDQ6VXNlcjMyNTU5OTQ= | User | false |
huggingface/diffusers | 4,652,167,775 | I_kwDOHa8MBc8AAAABFUpyXw | 13,930 | https://github.com/huggingface/diffusers/issues/13930 | https://api.github.com/repos/huggingface/diffusers/issues/13930 | LTX2 text connectors pass reversed prompt tokens and misplaced registers to the transformer (regression from #13564) | ### Describe the bug
`LTX2ConnectorTransformer1d.forward` (used by both the LTX 2.0 and LTX 2.3 pipelines) lays out its input sequence incorrectly before running the connector blocks: the prompt tokens reach the transformer **in reversed order**, and the learnable registers that fill the rest of the sequence are place... | closed | completed | false | 0 | [
"bug",
"models",
"pipelines"
] | [] | 2026-06-12T19:51:42Z | 2026-06-19T02:56:26Z | 2026-06-19T02:56:26Z | CONTRIBUTOR | null | 20260619T033516Z | 2026-06-19T03:35:16Z | Boffee | 8,808,565 | MDQ6VXNlcjg4MDg1NjU= | User | false |
huggingface/diffusers | 4,655,615,350 | I_kwDOHa8MBc8AAAABFX8Ndg | 13,935 | https://github.com/huggingface/diffusers/issues/13935 | https://api.github.com/repos/huggingface/diffusers/issues/13935 | Please reconsider `transformers>=5.0.0` | ### Describe the bug
Can you please reconsider requiring `transformers>=5.0.0` on the module-level solely for the optional prompt-enhancer (`pe`) component?
This is a significant burden for downstream projects that pin `transformers` 4.x for unrelated reasons. As one data point, migrating to `transformers` v5 in our... | open | null | false | 3 | [
"bug"
] | [] | 2026-06-13T13:15:46Z | 2026-06-16T00:26:29Z | null | CONTRIBUTOR | null | 20260616T003521Z | 2026-06-16T00:35:21Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,668,541,116 | I_kwDOHa8MBc8AAAABFkRIvA | 13,967 | https://github.com/huggingface/diffusers/issues/13967 | https://api.github.com/repos/huggingface/diffusers/issues/13967 | Can't use `load_lora_weights()` on `ErnieImageModularPipeline` | Hello,
When I try to load a LoRA with `load_lora_weights()` directly on a modular ERNIE-Image pipeline, library crashes saying such method doesn't exist.
PS: I submitted a PR before creating this issue.
### Reproduction
```python
manager = ComponentsManager()
pipe = ModularPipeline.from_pretrained(
"baidu/ERNI... | closed | completed | false | 0 | [
"bug",
"lora",
"modular-pipelines"
] | [] | 2026-06-15T20:11:23Z | 2026-06-16T00:53:57Z | 2026-06-16T00:53:57Z | CONTRIBUTOR | null | 20260616T013520Z | 2026-06-16T01:35:20Z | SamuelTallet | 36,248,671 | MDQ6VXNlcjM2MjQ4Njcx | User | false |
huggingface/diffusers | 4,686,820,159 | I_kwDOHa8MBc8AAAABF1szPw | 13,979 | https://github.com/huggingface/diffusers/issues/13979 | https://api.github.com/repos/huggingface/diffusers/issues/13979 | Can't use `load_lora_weights()` on `Ideogram4ModularPipeline` | Hello,
When I try to load a LoRA with `load_lora_weights()` directly on a modular Ideogram 4 pipeline, library crashes saying such method doesn't exist.
### Reproduction
```python
manager = ComponentsManager()
pipe = ModularPipeline.from_pretrained(
"CalamitousFelicitousness/Ideogram-4-bf16-Diffusers",
comp... | closed | completed | false | 0 | [
"bug",
"lora",
"modular-pipelines"
] | [] | 2026-06-17T20:26:23Z | 2026-06-17T22:37:12Z | 2026-06-17T22:37:12Z | CONTRIBUTOR | null | 20260617T233521Z | 2026-06-17T23:35:21Z | SamuelTallet | 36,248,671 | MDQ6VXNlcjM2MjQ4Njcx | User | false |
huggingface/diffusers | 4,691,827,083 | I_kwDOHa8MBc8AAAABF6eZiw | 13,984 | https://github.com/huggingface/diffusers/issues/13984 | https://api.github.com/repos/huggingface/diffusers/issues/13984 | FLUX LoRA with CLIP text-encoder weights fails (empty rank -> IndexError) under transformers>=5 | ### Describe the bug
Loading a kohya-style FLUX LoRA that contains **CLIP text-encoder** weights (`lora_te1_*`) fails with `IndexError: list index out of range` when `transformers>=5` is installed.
The text-encoder half of the LoRA is never loaded; `pipe.load_lora_weights(...)` raises before any adapter is injected.
... | open | null | false | 3 | [
"bug",
"lora",
"pipelines"
] | [] | 2026-06-18T11:16:30Z | 2026-06-21T20:11:41Z | null | CONTRIBUTOR | null | 20260621T203513Z | 2026-06-21T20:35:13Z | christopher5106 | 6,875,375 | MDQ6VXNlcjY4NzUzNzU= | User | false |
huggingface/diffusers | 4,699,583,637 | I_kwDOHa8MBc8AAAABGB30lQ | 13,998 | https://github.com/huggingface/diffusers/issues/13998 | https://api.github.com/repos/huggingface/diffusers/issues/13998 | FLUX kohya LoRA conversion crashes on `final_layer` (KeyError on missing adaLN_modulation_1; "Incompatible keys" on final_layer alphas) | ### Describe the bug
The kohya / sd-scripts FLUX LoRA converter (`_convert_sd_scripts_to_ai_toolkit` in `lora_conversion_utils.py`) crashes on LoRAs that include `final_layer` weights, in two distinct ways. Both reproduce on `diffusers==0.38.0` and on current `main`.
**Case A — `KeyError` when `final_layer.linear` ha... | closed | completed | false | 3 | [
"bug",
"lora",
"pipelines"
] | [] | 2026-06-19T09:30:12Z | 2026-06-20T08:47:12Z | 2026-06-20T08:47:12Z | CONTRIBUTOR | null | 20260620T093512Z | 2026-06-20T09:35:12Z | christopher5106 | 6,875,375 | MDQ6VXNlcjY4NzUzNzU= | User | false |
huggingface/diffusers | 4,701,347,353 | I_kwDOHa8MBc8AAAABGDjeGQ | 13,999 | https://github.com/huggingface/diffusers/issues/13999 | https://api.github.com/repos/huggingface/diffusers/issues/13999 | Flash Attention 3 varlen wrapper assumes output is a tuple | ### Describe the bug
`_flash_attention_3_varlen_hub` assumes `flash_attn_varlen_func` returns a tuple and does:
```py
out, lse, *_ = func(...)
```
The hub FA3 varlen function however, returns a single tensor with shape `[total_q, heads, dim]`, so Python destructuring takes the first sequence row as out.
### Repr... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-06-19T14:11:05Z | 2026-07-02T16:47:51Z | 2026-07-02T16:47:51Z | CONTRIBUTOR | null | 20260702T173518Z | 2026-07-02T17:35:18Z | bghira | 59,658,056 | MDQ6VXNlcjU5NjU4MDU2 | User | false |
huggingface/diffusers | 4,705,853,873 | I_kwDOHa8MBc8AAAABGH2hsQ | 14,002 | https://github.com/huggingface/diffusers/issues/14002 | https://api.github.com/repos/huggingface/diffusers/issues/14002 | Support Original Checkpoint-Compatible PEFT Adapters | **Is your feature request related to a problem? Please describe.**
My understanding is that `diffusers` does not currently support original checkpoint-compatible PEFT adapters aside from LoRAs.
**Describe the solution you'd like.**
I think `diffusers` should support loading general `peft`-compatible checkpoints which ... | open | null | false | 1 | [] | [] | 2026-06-20T09:03:09Z | 2026-07-01T19:48:18Z | null | MEMBER | null | 20260701T203521Z | 2026-07-01T20:35:21Z | dg845 | 58,458,699 | MDQ6VXNlcjU4NDU4Njk5 | User | false |
huggingface/diffusers | 4,705,938,472 | I_kwDOHa8MBc8AAAABGH7sKA | 14,003 | https://github.com/huggingface/diffusers/issues/14003 | https://api.github.com/repos/huggingface/diffusers/issues/14003 | Split Q,K,V vs Fused QKV Projections in diffusers Discussion | `diffusers` generally splits fused QKV projections to separate Q, K, V projections in its checkpoint format (for example, for diffusion transformer models). I am interested in collecting feedback on this design choice and any problems which may arise from it. To be honest upfront, I think it is unlikely that we will ch... | open | null | false | 3 | [] | [] | 2026-06-20T09:28:45Z | 2026-06-23T02:03:57Z | null | MEMBER | null | 20260623T023515Z | 2026-06-23T02:35:15Z | dg845 | 58,458,699 | MDQ6VXNlcjU4NDU4Njk5 | User | false |
huggingface/diffusers | 4,706,040,306 | I_kwDOHa8MBc8AAAABGIB58g | 14,004 | https://github.com/huggingface/diffusers/issues/14004 | https://api.github.com/repos/huggingface/diffusers/issues/14004 | PixArtSigmaPipeline: no LoRA loading support | ### Describe the bug
~~Calling `pipe.load_lora_weights(...)` on `ErnieImagePipeline` raises `AttributeError: 'ErnieImagePipeline' object has no attribute 'load_lora_weights'`. Neither the pipeline nor its denoiser, `ErnieImageTransformer2DModel`, currently has LoRA support wired up:~~
- ~~`ErnieImagePipeline` doesn't... | open | null | false | 1 | [] | [] | 2026-06-20T10:00:00Z | 2026-06-29T10:35:32Z | null | CONTRIBUTOR | null | 20260629T113512Z | 2026-06-29T11:35:12Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,706,075,671 | I_kwDOHa8MBc8AAAABGIEEFw | 14,005 | https://github.com/huggingface/diffusers/issues/14005 | https://api.github.com/repos/huggingface/diffusers/issues/14005 | _convert_unet_lora_key misses conv_in/conv_out/time_embedding mapping for Kohya UNet LoRAs | Note: I've been running some automated LoRA loading tests here https://github.com/Nerogar/OneTrainer/issues/1544
Many issues on OneTrainer's side to fix, but where I think it's a diffusers issue I'm reporting it upstream, but these can be about old or even obsolete features such as support Kohya-format LoRAs for SD 1.5... | closed | completed | false | 0 | [] | [] | 2026-06-20T10:09:56Z | 2026-06-28T15:54:44Z | 2026-06-28T15:54:44Z | CONTRIBUTOR | null | 20260628T163513Z | 2026-06-28T16:35:13Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,706,330,298 | I_kwDOHa8MBc8AAAABGITmug | 14,008 | https://github.com/huggingface/diffusers/issues/14008 | https://api.github.com/repos/huggingface/diffusers/issues/14008 | from_pretrained fails loading deprecated pipelines | Note: I understand these pipelines are deprecated and bugfixes aren't accepted anymore. But this seems to be a bug in _the process of moving pipelines to deprecated itself_
### Describe the bug
`_get_custom_components_and_folders` (`src/diffusers/pipelines/pipeline_loading_utils.py`) resolves custom pipeline compone... | open | null | false | 1 | [] | [] | 2026-06-20T11:23:45Z | 2026-06-21T15:44:05Z | null | CONTRIBUTOR | null | 20260621T163512Z | 2026-06-21T16:35:12Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,706,407,523 | I_kwDOHa8MBc8AAAABGIYUYw | 14,009 | https://github.com/huggingface/diffusers/issues/14009 | https://api.github.com/repos/huggingface/diffusers/issues/14009 | Kohya->diffusers LoRA key converter fails for Chroma and Flux1, if LoRA is not attention-only | **Describe the bug**
[`_convert_mixture_state_dict_to_diffusers`](https://github.com/huggingface/diffusers/blob/0f1abc4ae8b0eb2a3b40e82a310507281144c423/src/diffusers/loaders/lora_conversion_utils.py#L674) converts flat Kohya-style keys like `lora_transformer_transformer_blocks_0_<sublayer>` into dotted diffusers path... | closed | completed | false | 1 | [] | [] | 2026-06-20T11:46:56Z | 2026-06-27T10:32:59Z | 2026-06-27T10:32:59Z | CONTRIBUTOR | null | 20260627T103513Z | 2026-06-27T10:35:13Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,707,239,089 | I_kwDOHa8MBc8AAAABGJLEsQ | 14,012 | https://github.com/huggingface/diffusers/issues/14012 | https://api.github.com/repos/huggingface/diffusers/issues/14012 | attention dispatcher assumes wrong attributes for flash attn kernel from hub | ### Describe the bug
The breakage was in the active diffusers checkout, not in kernels. kernels-community/flash-attn2 exposes:
- `flash_attn_varlen_func`
- `flash_attn_varlen_qkvpacked_func`
- `flash_attn_interface._flash_attn_varlen_forward`
- `flash_attn_interface._flash_attn_varlen_backward`
But diffusers was t... | open | null | false | 0 | [
"bug"
] | [] | 2026-06-20T15:18:03Z | 2026-06-20T15:18:16Z | null | CONTRIBUTOR | null | 20260620T153512Z | 2026-06-20T15:35:12Z | bghira | 59,658,056 | MDQ6VXNlcjU5NjU4MDU2 | User | false |
huggingface/diffusers | 4,708,729,571 | I_kwDOHa8MBc8AAAABGKmC4w | 14,013 | https://github.com/huggingface/diffusers/issues/14013 | https://api.github.com/repos/huggingface/diffusers/issues/14013 | A bug on the set_timesteps function of the FlowMatchEulerDiscreteScheduler | ### Describe the bug
We fix a bug in the FlowMatchEulerDiscreteScheduler. If we call set_timesteps(num_inference_steps=2, timesteps=[1000. , 2.99401209]), the timesteps will be different from when we call set_timesteps(num_inference_steps=2), when the shift factor is not 1. However, its sigma list is equal.... | closed | completed | false | 6 | [
"bug"
] | [] | 2026-06-20T23:45:37Z | 2026-07-07T08:45:58Z | 2026-07-07T08:45:58Z | CONTRIBUTOR | null | 20260707T093517Z | 2026-07-07T09:35:17Z | TheLovesOfLadyPurple | 227,429,101 | U_kgDODY5K7Q | User | false |
huggingface/diffusers | 4,710,872,402 | I_kwDOHa8MBc8AAAABGMo1Ug | 14,025 | https://github.com/huggingface/diffusers/issues/14025 | https://api.github.com/repos/huggingface/diffusers/issues/14025 | MagCache on Wan 2.2 Dual-Transformer Pipelines: Incorrect Step Accounting and Limited Effectiveness on a 4-Step Distilled Model (e.g. Wan2.2) |
</h2><p>While evaluating MagCache on Wan 2.2, I found what appears to be a design issue in the current implementation for dual-transformer pipelines.</p><p>The current implementation assumes that a transformer executes exactly <code inline="">num_inference_steps</code> times during inference. This assumption is valid ... | open | null | false | 3 | [] | [] | 2026-06-21T14:49:14Z | 2026-07-03T07:12:25Z | null | NONE | null | 20260703T073519Z | 2026-07-03T07:35:19Z | fatemehgh1313 | 49,581,451 | MDQ6VXNlcjQ5NTgxNDUx | User | false |
huggingface/diffusers | 4,714,353,979 | I_kwDOHa8MBc8AAAABGP9VOw | 14,033 | https://github.com/huggingface/diffusers/issues/14033 | https://api.github.com/repos/huggingface/diffusers/issues/14033 | Remove link to HEIM for diffusion model evaluation | The doc page https://huggingface.co/docs/diffusers/conceptual/evaluation currently suggests using [HEIM](https://crfm.stanford.edu/helm/heim/latest/) for diffusion model evaluation.
Unfortunately, HEIM is no longer officially supported or actively maintained (see stanford-crfm/helm#2161, stanford-crfm/helm#3741) and is... | open | null | false | 0 | [] | [] | 2026-06-22T07:16:40Z | 2026-06-22T07:16:40Z | null | NONE | null | 20260622T073552Z | 2026-06-22T07:35:52Z | beyandbey | 231,219,854 | U_kgDODcgijg | User | false |
huggingface/diffusers | 4,715,727,756 | I_kwDOHa8MBc8AAAABGRRLjA | 14,037 | https://github.com/huggingface/diffusers/issues/14037 | https://api.github.com/repos/huggingface/diffusers/issues/14037 | `HookRegistry._child_registries_cache` goes stale after `enable_cache`/`disable_cache` → `ValueError: No context is set` | ### Describe the bug
`HookRegistry._get_child_registries()` caches the list of child-module registries (built by walking `named_modules()`) in `self._child_registries_cache` on first use and **never invalidates it**. But `register_hook` / `remove_hook` — invoked by `enable_cache()` / `disable_cache()` — change *which*... | open | null | false | 1 | [] | [] | 2026-06-22T10:29:33Z | 2026-06-29T22:23:31Z | null | CONTRIBUTOR | null | 20260629T223515Z | 2026-06-29T22:35:15Z | christopher5106 | 6,875,375 | MDQ6VXNlcjY4NzUzNzU= | User | false |
huggingface/diffusers | 4,718,647,167 | I_kwDOHa8MBc8AAAABGUDXfw | 14,041 | https://github.com/huggingface/diffusers/issues/14041 | https://api.github.com/repos/huggingface/diffusers/issues/14041 | Finalize Guidelines for Ulysses based context parallel attention | The PR https://github.com/huggingface/diffusers/pull/13924 added support for Ulysses Context parallel attention but looks like Library wide support for such an attention pattern still needs to be decided. | open | null | false | 0 | [] | [] | 2026-06-22T16:35:20Z | 2026-06-22T17:14:15Z | null | CONTRIBUTOR | null | 20260622T173518Z | 2026-06-22T17:35:18Z | atharvajoshi10 | 30,052,998 | MDQ6VXNlcjMwMDUyOTk4 | User | false |
huggingface/diffusers | 4,719,308,399 | I_kwDOHa8MBc8AAAABGUrubw | 14,042 | https://github.com/huggingface/diffusers/issues/14042 | https://api.github.com/repos/huggingface/diffusers/issues/14042 | Add Flux2KleinImg2ImgPipeline for FLUX.2 Klein | **Is your feature request related to a problem? Please describe.**
Currently, the `diffusers` library supports the base text-to-image `Flux2KleinPipeline`, but it lacks native support for Image-to-Image generation (`Img2Img`) for the FLUX.2 Klein architecture. Because Klein uses a highly optimized, compact structure (s... | closed | completed | false | 2 | [] | [] | 2026-06-22T18:05:17Z | 2026-06-25T11:18:47Z | 2026-06-25T11:18:47Z | NONE | null | 20260625T113518Z | 2026-06-25T11:35:18Z | Sainava | 158,836,146 | U_kgDOCXelsg | User | false |
huggingface/diffusers | 4,722,190,618 | I_kwDOHa8MBc8AAAABGXbpGg | 14,049 | https://github.com/huggingface/diffusers/issues/14049 | https://api.github.com/repos/huggingface/diffusers/issues/14049 | Add JoyAI-Image Edit Plus pipeline and model | ### Model/Pipeline/Scheduler description
## Model description
__JoyAI-Image__ is a unified multimodal foundation model for image understanding, text-to-image generation, and instruction-guided image editing from JD.com. It combines an 8B
Multimodal Large Language Model (MLLM) with a 16B Multimodal Diffus... | closed | completed | false | 0 | [] | [] | 2026-06-23T02:52:02Z | 2026-07-08T00:45:08Z | 2026-07-08T00:45:08Z | CONTRIBUTOR | null | 20260708T013512Z | 2026-07-08T01:35:12Z | tangyanf | 27,288,350 | MDQ6VXNlcjI3Mjg4MzUw | User | false |
huggingface/diffusers | 4,727,746,885 | I_kwDOHa8MBc8AAAABGcuxRQ | 14,052 | https://github.com/huggingface/diffusers/issues/14052 | https://api.github.com/repos/huggingface/diffusers/issues/14052 | Fix mutable default argument in lora_base.py | **Issue:** Found mutable default arguments (`[]`, `{}`) in lora_base.py and related files. This is a classic Python bug pattern.
**Problem:** Mutable default arguments are created once at function definition time, and shared across all calls. This can cause unexpected state leaks between function calls.
**Suggested f... | closed | completed | false | 1 | [] | [] | 2026-06-23T16:37:33Z | 2026-07-02T13:06:55Z | 2026-07-02T13:06:55Z | NONE | null | 20260702T133519Z | 2026-07-02T13:35:19Z | shunfeng8421 | 37,066,911 | MDQ6VXNlcjM3MDY2OTEx | User | false |
huggingface/diffusers | 4,729,625,305 | I_kwDOHa8MBc8AAAABGeha2Q | 14,056 | https://github.com/huggingface/diffusers/issues/14056 | https://api.github.com/repos/huggingface/diffusers/issues/14056 | Modular pipeline: Krea 2 | Hello, thanks for your great work!
Could you please implement `Krea2ModularPipeline`?
@Abhinay1997 @EleaZhong | open | null | false | 3 | [
"good first issue",
"contributions-welcome"
] | [] | 2026-06-23T21:20:15Z | 2026-06-29T00:00:38Z | null | CONTRIBUTOR | null | 20260629T003513Z | 2026-06-29T00:35:13Z | SamuelTallet | 36,248,671 | MDQ6VXNlcjM2MjQ4Njcx | User | false |
huggingface/diffusers | 4,729,648,827 | I_kwDOHa8MBc8AAAABGei2uw | 14,057 | https://github.com/huggingface/diffusers/issues/14057 | https://api.github.com/repos/huggingface/diffusers/issues/14057 | SD1.5 - pipeline_controlnet_img2img and pipeline_controlnet_inpaint are mixing variables "image" and "control_image" | ### Describe the bug
"image" is used in pipeline_controlnet for a list of controlnet images.
"control_image" is used in pipeline_controlnet_img2img and pipeline_controlnet_inpaint for a list of controlnet images.
img2img and inpaint are missing "control_image" in check_inputs function causing an error
The img2img and... | open | null | false | 1 | [
"bug"
] | [] | 2026-06-23T21:24:32Z | 2026-06-23T22:12:41Z | null | NONE | null | 20260623T223518Z | 2026-06-23T22:35:18Z | buddyforgothisemail | 296,296,018 | U_kgDOEakeUg | User | false |
huggingface/diffusers | 4,734,631,494 | I_kwDOHa8MBc8AAAABGjS-Rg | 14,060 | https://github.com/huggingface/diffusers/issues/14060 | https://api.github.com/repos/huggingface/diffusers/issues/14060 | Support for LoRA / adapters in AceStepPipeline | Hello,
First of all, thank you for adding the ACE-Step 1.5 pipeline. It's working nicely!
I noticed that LoRA support doesn't seem to be implemented yet in _**AceStepPipeline**_.
When trying to load a LoRA I get this error:
_**'AceStepPipeline' object has no attribute 'set_adapters'**_
(using the standard Diffusers met... | open | null | false | 2 | [] | [] | 2026-06-24T12:05:49Z | 2026-06-24T17:26:38Z | null | NONE | null | 20260624T173512Z | 2026-06-24T17:35:12Z | Nib-git | 295,265,852 | U_kgDOEZlmPA | User | false |
huggingface/diffusers | 4,741,672,390 | I_kwDOHa8MBc8AAAABGqAtxg | 14,063 | https://github.com/huggingface/diffusers/issues/14063 | https://api.github.com/repos/huggingface/diffusers/issues/14063 | Kandinsky5 pipeline does not support device_map=balanced | ### Describe the bug
Kandinsky5 I2I/I2V failures with multi-GPU device_map: Under device_map="balanced", VAE-encoded image latents may be produced on a different GPU than the generated noise latents. And it will return error:
`RuntimeError: Expected all tensors to be on the same device, but got tensors is on cuda:1, d... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-06-25T07:25:46Z | 2026-06-30T06:06:23Z | 2026-06-30T06:06:23Z | CONTRIBUTOR | null | 20260630T063512Z | 2026-06-30T06:35:12Z | kaixuanliu | 13,268,042 | MDQ6VXNlcjEzMjY4MDQy | User | false |
huggingface/diffusers | 4,742,948,631 | I_kwDOHa8MBc8AAAABGrOnFw | 14,065 | https://github.com/huggingface/diffusers/issues/14065 | https://api.github.com/repos/huggingface/diffusers/issues/14065 | DreamLite-base fails to load due to unsupported legacy UNet block type names | `DreamLitePipeline.from_pretrained("carlofkl/DreamLite-base")` currently fails while constructing the UNet because the saved config uses legacy/upstream-style block type names that `DreamLiteUNetModel` does not accept.
Minimal repro:
```python
from diffusers import DreamLitePipeline
pipe = DreamLitePipeline.f... | closed | completed | false | 0 | [] | [] | 2026-06-25T10:27:43Z | 2026-06-26T23:42:20Z | 2026-06-26T23:42:20Z | CONTRIBUTOR | null | 20260627T003512Z | 2026-06-27T00:35:12Z | ElectricGoal | 40,719,019 | MDQ6VXNlcjQwNzE5MDE5 | User | false |
huggingface/diffusers | 4,745,704,098 | I_kwDOHa8MBc8AAAABGt2yog | 14,067 | https://github.com/huggingface/diffusers/issues/14067 | https://api.github.com/repos/huggingface/diffusers/issues/14067 | uv pip install -e ".[test]" fails on Python 3.12 due to incompatible numba/llvmlite resolution | ### Describe the bug
Installing the test dependencies from a fresh clone fails on Python 3.12.
`librosa==0.11.0` permits `numba>=0.51.0`, and `uv` resolves this to `numba==0.53.1`, which requires `llvmlite==0.36.0`. That version of `llvmlite` only supports Python `<3.10`.
This occurs with the current `uv==0.11.24` a... | closed | completed | false | 0 | [] | [] | 2026-06-25T16:36:54Z | 2026-07-03T02:34:30Z | 2026-07-03T02:34:30Z | NONE | null | 20260703T023518Z | 2026-07-03T02:35:18Z | gabe-engineers | 39,689,869 | MDQ6VXNlcjM5Njg5ODY5 | User | false |
huggingface/diffusers | 4,750,779,229 | I_kwDOHa8MBc8AAAABGysjXQ | 14,072 | https://github.com/huggingface/diffusers/issues/14072 | https://api.github.com/repos/huggingface/diffusers/issues/14072 | [Bug] Layerwise casting tests fail on MPS backend due to float8 unsupported dtype | ### Describe the bug
When running the unit test suite on Apple Silicon (macOS) using the MPS backend (`torch_device = "mps"`), tests checking `enable_layerwise_casting` fail. This occurs because the MPS backend does not support the `torch.float8_e4m3fn` dtype.
Several test cases unconditionally configure layerwise ca... | closed | completed | false | 0 | [] | [] | 2026-06-26T09:09:23Z | 2026-06-29T13:38:02Z | 2026-06-29T13:38:02Z | CONTRIBUTOR | null | 20260629T143526Z | 2026-06-29T14:35:26Z | GiGiKoneti | 207,724,134 | U_kgDODGGeZg | User | false |
huggingface/diffusers | 4,755,220,009 | I_kwDOHa8MBc8AAAABG27mKQ | 14,076 | https://github.com/huggingface/diffusers/issues/14076 | https://api.github.com/repos/huggingface/diffusers/issues/14076 | [Proposal]: Support LoRA loading for MotifVideo pipelines | ## Summary
I noticed that the recently added MotifVideo pipelines currently do not expose pipeline-level LoRA loading support, even though the underlying transformer appears to already be PEFT-ready.
Before working on a PR, I'd like to check whether adding LoRA support for the MotifVideo pipelines would be considered... | closed | completed | false | 5 | [] | [] | 2026-06-26T21:21:14Z | 2026-07-09T08:02:40Z | 2026-07-09T08:02:40Z | NONE | null | 20260709T083518Z | 2026-07-09T08:35:18Z | Sainava | 158,836,146 | U_kgDOCXelsg | User | false |
huggingface/diffusers | 4,766,540,312 | I_kwDOHa8MBc8AAAABHBuiGA | 14,085 | https://github.com/huggingface/diffusers/issues/14085 | https://api.github.com/repos/huggingface/diffusers/issues/14085 | Backport workflow-hardening fix (excessive-permissions) to 3 release branches | ### Summary
The default branch already hardened `.github/workflows/stale.yml` against the issue(s) below, but **3** release branches still carry it. This proposes the same, minimal, scanner-verified fix for each.
### What's flagged (by [zizmor](https://github.com/woodruffw/zizmor))
- `excessive-permissions` — workflow... | open | null | false | 0 | [] | [] | 2026-06-29T08:52:11Z | 2026-06-29T08:52:11Z | null | NONE | null | 20260629T093513Z | 2026-06-29T09:35:13Z | CharlieMCY | 26,402,240 | MDQ6VXNlcjI2NDAyMjQw | User | false |
huggingface/diffusers | 4,766,581,529 | I_kwDOHa8MBc8AAAABHBxDGQ | 14,086 | https://github.com/huggingface/diffusers/issues/14086 | https://api.github.com/repos/huggingface/diffusers/issues/14086 | Add SeFi-Image pipeline | ### Model/Pipeline/Scheduler description
We are proposing the integration of SeFi-Image into Diffusers.
SeFi-Image is a text-to-image model family built with Semantic-First Diffusion. It separates image generation into semantic and texture latent streams, denoising the semantic structure slightly ahead of texture det... | open | null | false | 0 | [] | [] | 2026-06-29T08:58:15Z | 2026-06-29T08:58:15Z | null | NONE | null | 20260629T093513Z | 2026-06-29T09:35:13Z | jmliu206 | 57,990,373 | MDQ6VXNlcjU3OTkwMzcz | User | false |
huggingface/diffusers | 4,767,382,464 | I_kwDOHa8MBc8AAAABHCh7wA | 14,088 | https://github.com/huggingface/diffusers/issues/14088 | https://api.github.com/repos/huggingface/diffusers/issues/14088 | Handling of .alpha in from-comfy converters | ### Describe the bug
This is a question and maybe a bug report:
LoRAs in diffusers/peft-format don't support an `alpha` scaling factor. when converting another format to diffusers/peft, .alpha can be folded into lora_B]
kohya format LoRAs do hgave .alpha and for those models that diffusers implement a from-kohya conv... | open | null | false | 8 | [
"bug"
] | [] | 2026-06-29T10:45:46Z | 2026-06-29T21:28:12Z | null | CONTRIBUTOR | null | 20260629T213515Z | 2026-06-29T21:35:15Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,768,945,636 | I_kwDOHa8MBc8AAAABHEBV5A | 14,090 | https://github.com/huggingface/diffusers/issues/14090 | https://api.github.com/repos/huggingface/diffusers/issues/14090 | Fix model offloading and training tests + prevent examples timeout | ### Describe the bug
When running the full test suite on models and training examples, three distinct test failures/hangs consistently block test runners:
1. **AutoencoderVidTok return format mismatch:**
`AutoencoderVidTok.forward` returns a raw tensor `dec` when `return_dict=False`. However, the standard VAE/auto... | closed | completed | false | 0 | [] | [] | 2026-06-29T14:20:56Z | 2026-07-04T03:02:41Z | 2026-07-04T03:02:41Z | CONTRIBUTOR | null | 20260704T033517Z | 2026-07-04T03:35:17Z | GiGiKoneti | 207,724,134 | U_kgDODGGeZg | User | false |
huggingface/diffusers | 4,773,093,287 | I_kwDOHa8MBc8AAAABHH-fpw | 14,094 | https://github.com/huggingface/diffusers/issues/14094 | https://api.github.com/repos/huggingface/diffusers/issues/14094 | Reduce prompt costs with SuperCompress compression | ## SuperCompress: Open-Source Prompt Compression
[SuperCompress](https://github.com/arjunkshah/supercompress) is a learned prompt compression library that reduces LLM token usage by ~65% using a ~5K parameter neural policy.
**Why this matters for ML development:**
- Practical ML deployment: a tiny model (~5K params) ... | closed | completed | false | 1 | [] | [] | 2026-06-30T02:17:30Z | 2026-06-30T04:06:53Z | 2026-06-30T04:06:53Z | NONE | null | 20260630T043512Z | 2026-06-30T04:35:12Z | arjunkshah | 180,205,446 | U_kgDOCr23hg | User | false |
huggingface/diffusers | 4,773,154,992 | I_kwDOHa8MBc8AAAABHICQsA | 14,095 | https://github.com/huggingface/diffusers/issues/14095 | https://api.github.com/repos/huggingface/diffusers/issues/14095 | SuperCompress: Free open-source LLM prompt compression - cut token costs by ~65% | ## SuperCompress: Open-Source Prompt Compression
[SuperCompress](https://github.com/arjunkshah/supercompress) is a learned prompt compression library that reduces LLM token usage by ~65% using a ~5K parameter neural policy.
**Why this matters for ML development:**
- Practical ML deployment: a tiny model (~5K params) ... | closed | completed | false | 0 | [] | [] | 2026-06-30T02:32:12Z | 2026-06-30T04:17:28Z | 2026-06-30T04:17:28Z | NONE | null | 20260630T043512Z | 2026-06-30T04:35:12Z | arjunkshah | 180,205,446 | U_kgDOCr23hg | User | false |
huggingface/diffusers | 4,781,564,832 | I_kwDOHa8MBc8AAAABHQDjoA | 14,097 | https://github.com/huggingface/diffusers/issues/14097 | https://api.github.com/repos/huggingface/diffusers/issues/14097 | Regression (#13485) Broken TorchAO Compat | ### Describe the bug
Some changes introduced in https://github.com/huggingface/diffusers/pull/13485 / https://github.com/huggingface/diffusers/commit/236e5dd9f38e21ae40c002539368b9be9a5e0fc8 have regressed compat with compiled 8-bit models through TorchAO.
### Reproduction
Simply `./regression.py`. Tested on 1xH200.... | open | null | false | 3 | [
"bug"
] | [] | 2026-07-01T01:19:41Z | 2026-07-02T09:15:02Z | null | CONTRIBUTOR | null | 20260702T093518Z | 2026-07-02T09:35:18Z | Beinsezii | 39,478,211 | MDQ6VXNlcjM5NDc4MjEx | User | false |
huggingface/diffusers | 4,788,367,084 | I_kwDOHa8MBc8AAAABHWiu7A | 14,103 | https://github.com/huggingface/diffusers/issues/14103 | https://api.github.com/repos/huggingface/diffusers/issues/14103 | Custom AttentionProcessor lost/fallback occurring after forward pass | ### Describe the bug
I am attempting to inject a custom attention processor (IPAdapter) into the UNet2DConditionModel attention blocks (specifically attn2 layers) using set_attn_processor. While the processor is correctly assigned and verified immediately after the forward passs it is being silently overwritten.
Upon... | closed | completed | false | 1 | [
"bug",
"models"
] | [] | 2026-07-01T18:50:03Z | 2026-07-01T19:20:01Z | 2026-07-01T19:14:40Z | NONE | null | 20260701T193518Z | 2026-07-01T19:35:18Z | abdullahejazjanjua | 113,930,351 | U_kgDOBspwbw | User | false |
huggingface/diffusers | 4,788,643,706 | I_kwDOHa8MBc8AAAABHWzneg | 14,104 | https://github.com/huggingface/diffusers/issues/14104 | https://api.github.com/repos/huggingface/diffusers/issues/14104 | Flux-family attention: flash-attn and other backends fail in an autocast context | ### Describe the bug
- q and k are normed before attention in flux-style attention.
- in an autocast context, norms are calculated in float32 no matter their input - and are returned as float32
- the result gets passed on to attention
- no problem for torch scaled_dot_product_attention, it can handle float32 - doesn't... | open | null | false | 6 | [
"bug"
] | [] | 2026-07-01T19:33:17Z | 2026-07-10T11:09:39Z | null | CONTRIBUTOR | null | 20260710T113512Z | 2026-07-10T11:35:12Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,801,449,498 | I_kwDOHa8MBc8AAAABHjBOGg | 14,114 | https://github.com/huggingface/diffusers/issues/14114 | https://api.github.com/repos/huggingface/diffusers/issues/14114 | `_flash_3_varlen_hub` backend cannot handle non-contiguous mask | ### Describe the bug
Model like Qwen-Image typically has a non-contiguous mask, currently `_flash_3_varlen_hub` backend does not support this, and cause numerical error in forward and backward silently.
### Reproduction
We and produce the bug using the following code snippet
```python
"""
Usage:
python test_fa3... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-07-03T07:42:46Z | 2026-07-06T05:03:13Z | 2026-07-06T05:03:13Z | CONTRIBUTOR | null | 20260706T053518Z | 2026-07-06T05:35:18Z | zhtmike | 8,342,575 | MDQ6VXNlcjgzNDI1NzU= | User | false |
huggingface/diffusers | 4,804,481,662 | I_kwDOHa8MBc8AAAABHl6Sfg | 14,117 | https://github.com/huggingface/diffusers/issues/14117 | https://api.github.com/repos/huggingface/diffusers/issues/14117 | `DiffusionPipeline.download()` breaks with huggingface_hub>=1.22.0 in offline mode (IncompleteSnapshotError) | ## Summary
`huggingface_hub` v1.22.0 introduced a breaking change: `snapshot_download(..., local_files_only=True)` now verifies that the cached snapshot is **complete** (it caches the `/tree` call, so even offline it can tell whether the cached folder is partial). Previously, in offline mode it returned whatever snaps... | open | null | false | 1 | [
"bug",
"needs-code-example",
"pipelines"
] | [] | 2026-07-03T15:03:01Z | 2026-07-03T15:06:09Z | null | NONE | null | 20260703T153517Z | 2026-07-03T15:35:17Z | moon-bot-app[bot] | 272,242,342 | BOT_kgDOEDoWpg | Bot | false |
huggingface/diffusers | 4,812,336,578 | I_kwDOHa8MBc8AAAABHtZtwg | 14,122 | https://github.com/huggingface/diffusers/issues/14122 | https://api.github.com/repos/huggingface/diffusers/issues/14122 | `from_single_file()` support missing/incomplete for Ideogram4 and Krea2 transformers | ### Describe the bug
Neither `Ideogram4Transformer2DModel` nor `Krea2Transformer2DModel` currently support `from_single_file()`, though
for different reasons.
**Ideogram4Transformer2DModel** inherits `FromOriginalModelMixin` (since it was added in #13859), but it was never
added to `SINGLE_FILE_LOADABLE_CLASSES` in... | open | null | false | 1 | [
"bug",
"models",
"single-file",
"feature-request"
] | [] | 2026-07-05T09:37:21Z | 2026-07-06T01:20:28Z | null | CONTRIBUTOR | null | 20260706T013516Z | 2026-07-06T01:35:16Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/diffusers | 4,812,770,588 | I_kwDOHa8MBc8AAAABHt0NHA | 14,124 | https://github.com/huggingface/diffusers/issues/14124 | https://api.github.com/repos/huggingface/diffusers/issues/14124 | [Version 0.39.0]AttributeError: 'CLIPTextModel' object has no attribute 'text_model' WHEN SET CKIP_SKIP NOT NONE | ### Describe the bug
This bug should have been fixed, yet it still occurs in the latest version.
It does not trigger when `clip_skip` is set to None, but the issue appears whenever `clip_skip` has a non-None value.
### Reproduction
```
pipe = StableDiffusionPipeline.from_single_file(
pretrained_model_link_o... | open | null | false | 0 | [
"bug",
"pipelines"
] | [] | 2026-07-05T12:10:16Z | 2026-07-05T12:11:09Z | null | NONE | null | 20260705T123518Z | 2026-07-05T12:35:18Z | CodingForAGI | 141,250,660 | U_kgDOCGtQZA | User | false |
huggingface/diffusers | 4,823,522,658 | I_kwDOHa8MBc8AAAABH4EdYg | 14,132 | https://github.com/huggingface/diffusers/issues/14132 | https://api.github.com/repos/huggingface/diffusers/issues/14132 | [modular] ensure branch-specific input defaults | **What API design would you like to have changed or added to the library? Why?**
Currently, if a `AutoPipelineBlocks` contains 2 branch blocks; if both have the same input, one branch block declares a default while the other does not; both blocks will have the same default, so that the default value set in one branch ... | open | null | false | 0 | [
"bug",
"modular-pipelines"
] | [
"yiyixuxu"
] | 2026-07-06T22:19:27Z | 2026-07-06T22:20:41Z | null | MEMBER | null | 20260706T223516Z | 2026-07-06T22:35:16Z | yiyixuxu | 12,631,849 | MDQ6VXNlcjEyNjMxODQ5 | User | false |
huggingface/diffusers | 4,831,406,120 | I_kwDOHa8MBc8AAAABH_loKA | 14,135 | https://github.com/huggingface/diffusers/issues/14135 | https://api.github.com/repos/huggingface/diffusers/issues/14135 | HookRegistry.register_hook breaks module forward signature | ### Describe the bug
Similar to #13803
HookRegistry.register_hook breaks a modules forward signature, and downstream prevents torch.compile / torch.export - specifically affecting context parallel hooks
### Reproduction
```
import inspect
import torch
from diffusers.hooks.hooks import HookRegistry, ModelHook
clas... | open | null | false | 0 | [
"bug"
] | [] | 2026-07-07T19:00:28Z | 2026-07-07T19:02:11Z | null | NONE | null | 20260707T193514Z | 2026-07-07T19:35:14Z | jloftin-nv | 288,774,625 | U_kgDOETZZ4Q | User | false |
huggingface/diffusers | 4,833,641,230 | I_kwDOHa8MBc8AAAABIBuDDg | 14,138 | https://github.com/huggingface/diffusers/issues/14138 | https://api.github.com/repos/huggingface/diffusers/issues/14138 | Boogu-Image | ### Model/Pipeline/Scheduler description
Does Diffusers support the Boogu-Image model?
### Open source status
- [x] The model implementation is available.
- [x] The model weights are available (Only relevant if addition is not a scheduler).
### Provide useful links for the implementation
https://huggingface.co/Boo... | open | null | false | 0 | [
"models",
"feature-request"
] | [] | 2026-07-08T01:56:05Z | 2026-07-08T01:56:41Z | null | NONE | null | 20260708T023512Z | 2026-07-08T02:35:12Z | JaDonghao | 182,836,634 | U_kgDOCuXdmg | User | false |
huggingface/diffusers | 4,833,799,264 | I_kwDOHa8MBc8AAAABIB3sYA | 14,139 | https://github.com/huggingface/diffusers/issues/14139 | https://api.github.com/repos/huggingface/diffusers/issues/14139 | Non-hub varlen attention backends mispack keys for non-contiguous masks (missed in #14114/#14115) | The three non-hub varlen backends pack keys/values by slicing `key[b, :valid_len]`, where `valid_len = attn_mask.sum(dim=1)` is only the *count* of valid keys. This assumes the valid keys are contiguous from index 0. For a left-padded or interior-gap mask, this packs the wrong keys: attention runs over padding position... | open | null | false | 2 | [
"bug",
"models"
] | [] | 2026-07-08T02:30:13Z | 2026-07-09T19:39:33Z | null | NONE | null | 20260709T203524Z | 2026-07-09T20:35:24Z | qflen | 194,738,340 | U_kgDOC5t4pA | User | false |
huggingface/diffusers | 4,836,734,216 | I_kwDOHa8MBc8AAAABIEq1CA | 14,142 | https://github.com/huggingface/diffusers/issues/14142 | https://api.github.com/repos/huggingface/diffusers/issues/14142 | DiffusionPipeline.download() on Windows silently skips all subfolder config.json files with huggingface_hub>=1.22.0 | ### Describe the bug
On Windows, `DiffusionPipeline.download()` builds the allow patterns for per-component configs with `os.path.join`:
https://github.com/huggingface/diffusers/blob/b8905b9b0f01/src/diffusers/pipelines/pipeline_utils.py#L1740-L1741
which produces `vae\config.json`, while hub repo paths always use for... | open | null | false | 2 | [
"bug",
"pipelines"
] | [] | 2026-07-08T10:48:14Z | 2026-07-09T19:41:37Z | null | CONTRIBUTOR | null | 20260709T203524Z | 2026-07-09T20:35:24Z | CalamitousFelicitousness | 63,501,758 | MDQ6VXNlcjYzNTAxNzU4 | User | false |
huggingface/diffusers | 4,841,905,986 | I_kwDOHa8MBc8AAAABIJmfQg | 14,145 | https://github.com/huggingface/diffusers/issues/14145 | https://api.github.com/repos/huggingface/diffusers/issues/14145 | `LTX2*Pipeline.from_single_file` cannot load LTX 2.3 (or 2.0) combined checkpoints: unmapped 2.3 keys leave meta tensors, and the LTX2 converters drain the shared checkpoint dict | ### Describe the bug
Pipeline-level `from_single_file` is broken for the Lightricks LTX-2 combined (flat) checkpoints. There are two independent bugs; both are present on current `main` (verified in source at commit `208704a`, 2026-07-08). **I intend to submit a pull request for this** — a minimal, verified fix is rea... | open | null | false | 1 | [
"bug",
"models",
"pipelines",
"single-file"
] | [] | 2026-07-08T23:40:46Z | 2026-07-09T19:22:34Z | null | NONE | null | 20260709T193523Z | 2026-07-09T19:35:23Z | cgu2022 | 44,711,033 | MDQ6VXNlcjQ0NzExMDMz | User | false |
huggingface/diffusers | 4,842,623,667 | I_kwDOHa8MBc8AAAABIKSSsw | 14,146 | https://github.com/huggingface/diffusers/issues/14146 | https://api.github.com/repos/huggingface/diffusers/issues/14146 | `Krea2Pipeline` produces incorrect results on ROCm/gfx1201 | ### Describe the bug
I used a gfx1201 GPU (Radeon RX 9070) to produce an image with Krea2, using torch 2.13.0, diffusers at commit 208704a27a6f362b67cd1a04fa1db0b98036d26f (recent `main`). It produces a garbled image. Every other pipeline tested works well.
<img width="1024" height="1024" alt="Image" src="https://git... | closed | completed | false | 5 | [
"bug",
"pipelines"
] | [] | 2026-07-09T02:25:50Z | 2026-07-12T20:19:03Z | 2026-07-09T13:56:15Z | CONTRIBUTOR | null | 20260712T203520Z | 2026-07-12T20:35:20Z | hameerabbasi | 2,190,658 | MDQ6VXNlcjIxOTA2NTg= | User | false |
huggingface/diffusers | 4,848,967,340 | I_kwDOHa8MBc8AAAABIQVerA | 14,149 | https://github.com/huggingface/diffusers/issues/14149 | https://api.github.com/repos/huggingface/diffusers/issues/14149 | Add AMD Quark support to Diffusers | **Is your feature request related to a problem? Please describe.**
Diffusers cannot load or run AMD Quark-quantized diffusion models through the standard from_pretrained APIs Today the only way to use a Quark-quantized UNet/transformer with Diffusers is to monkey-patch Quark into the auto-quantizer mappings at runtime,... | open | null | false | 0 | [
"quantization",
"models",
"feature-request"
] | [] | 2026-07-09T18:42:53Z | 2026-07-09T18:43:52Z | null | NONE | null | 20260709T193523Z | 2026-07-09T19:35:23Z | amd-inechakr | 257,418,406 | U_kgDOD1fkpg | User | false |
huggingface/diffusers | 4,867,656,143 | I_kwDOHa8MBc8AAAABIiKJzw | 14,175 | https://github.com/huggingface/diffusers/issues/14175 | https://api.github.com/repos/huggingface/diffusers/issues/14175 | Path traversal / arbitrary out-of-directory file read via sharded-checkpoint `weight_map` in `_get_checkpoint_shard_files` | ### Describe the bug
reported on 9 June 2026 https://github.com/huggingface/diffusers/security/advisories/GHSA-9mfc-5mqx-wxhr - no response.
## Summary
`DiffusionPipeline.from_pretrained` / `ModelMixin.from_pretrained` resolve a sharded checkpoint's shard filenames from the model-supplied index JSON (`diffusion_pyt... | open | null | false | 0 | [
"bug",
"needs-code-example",
"models",
"pipelines"
] | [] | 2026-07-12T15:40:16Z | 2026-07-12T15:41:08Z | null | NONE | null | 20260712T163520Z | 2026-07-12T16:35:20Z | geo-chen | 2,404,584 | MDQ6VXNlcjI0MDQ1ODQ= | User | false |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.