zhangyi Cursor commited on
Commit ·
bb76f59
1
Parent(s): c134b32
Update README affiliations, project page, and ICLR 2026 info.
Browse files
README.md
CHANGED
|
@@ -1,125 +1,130 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
library_name: diffusers
|
| 4 |
-
pipeline_tag: text-to-image
|
| 5 |
-
tags:
|
| 6 |
-
- flow-matching
|
| 7 |
-
- distillation
|
| 8 |
-
- flux
|
| 9 |
-
- stable-diffusion
|
| 10 |
-
---
|
| 11 |
-
|
| 12 |
-
# 🚀 SenseFlow: Scaling Distribution Matching for Flow-based Text-to-Image Distillation
|
| 13 |
-
|
| 14 |
-
[](https://arxiv.org/abs/2506.00523)
|
| 15 |
-
[](https://arxiv.org/abs/2506.00523)
|
| 15 |
+
[](https://xingtongge.github.io/SenseFlow-page)
|
| 16 |
+
[](https://github.com/XingtongGe/SenseFlow)
|
| 17 |
+
[](https://huggingface.co/domiso/SenseFlow)
|
| 18 |
+
|
| 19 |
+
[Xingtong Ge](https://xingtongge.github.io/)<sup>1,2</sup>, Xin Zhang<sup>3</sup>, [Tongda Xu](https://tongdaxu.github.io/)<sup>4</sup>, [Yi Zhang](https://zhangyi-3.github.io/)<sup>3</sup>, [Xinjie Zhang](https://xinjie-q.github.io/)<sup>1</sup>, [Yan Wang](https://yanwang202199.github.io/)<sup>4</sup>, [Jun Zhang](https://eejzhang.people.ust.hk/)<sup>1</sup><sup>*</sup>
|
| 20 |
+
|
| 21 |
+
<sup>1</sup>The Hong Kong University of Science and Technology, <sup>2</sup>SenseTime Research, <sup>3</sup>Vivix AI, <sup>4</sup>Institute for AI Industry Research, Tsinghua University
|
| 22 |
+
|
| 23 |
+
The Fourteenth International Conference on Learning Representations (**ICLR**), 2026
|
| 24 |
+
|
| 25 |
+
## Abstract
|
| 26 |
+
|
| 27 |
+
The Distribution Matching Distillation (DMD) has been successfully applied to text-to-image diffusion models such as Stable Diffusion (SD) 1.5. However, vanilla DMD suffers from convergence difficulties on large-scale flow-based text-to-image models, such as SD 3.5 and FLUX. In this paper, we first analyze the issues when applying vanilla DMD on large-scale models. Then, to overcome the scalability challenge, we propose implicit distribution alignment (IDA) to regularize the distance between the generator and fake distribution. Furthermore, we propose intra-segment guidance (ISG) to relocate the timestep importance distribution from the teacher model. With IDA alone, DMD converges for SD 3.5; employing both IDA and ISG, DMD converges for SD 3.5 and FLUX.1 dev. Along with other improvements such as scaled up discriminator models, our final model, dubbed **SenseFlow**, achieves superior performance in distillation for both diffusion based text-to-image models such as SDXL, and flow-matching models such as SD 3.5 Large and FLUX. The source code and model weights are now available.
|
| 28 |
+
|
| 29 |
+
## SenseFlow-FLUX.1 dev (supports 4–8-step generation)
|
| 30 |
+
* `SenseFlow-FLUX/diffusion_pytorch_model.safetensors`: the DiT checkpoint.
|
| 31 |
+
* `SenseFlow-FLUX/config.json`: the config of DiT using in our model.
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
### Usage
|
| 35 |
+
|
| 36 |
+
1. prepare the base checkpoint of FLUX.1 dev to `Path/to/FLUX`
|
| 37 |
+
2. Use `SenseFlow-FLUX` to replace the transformer folder `Path/to/FLUX/transformer`, obtaining the `Path/to/SenseFlow-FLUX`.
|
| 38 |
+
|
| 39 |
+
#### Using the Euler sampler
|
| 40 |
+
```python
|
| 41 |
+
import torch
|
| 42 |
+
from diffusers import FluxPipeline
|
| 43 |
+
from diffusers import FlowMatchEulerDiscreteScheduler
|
| 44 |
+
|
| 45 |
+
pipe = FluxPipeline.from_pretrained("Path/to/SenseFlow-FLUX", torch_dtype=torch.bfloat16).to("cuda")
|
| 46 |
+
|
| 47 |
+
prompt="A cat sleeping on a windowsill with white curtains fluttering in the breeze"
|
| 48 |
+
|
| 49 |
+
images = pipe(
|
| 50 |
+
prompt,
|
| 51 |
+
height=1024,
|
| 52 |
+
width=1024,
|
| 53 |
+
num_inference_steps=4,
|
| 54 |
+
max_sequence_length=512,
|
| 55 |
+
).images[0]
|
| 56 |
+
|
| 57 |
+
images.save("output.png")
|
| 58 |
+
```
|
| 59 |
+
#### Using the x0 sampler (similar to the LCMScheduler in diffusers)
|
| 60 |
+
```python
|
| 61 |
+
import torch
|
| 62 |
+
from diffusers import FluxPipeline
|
| 63 |
+
from diffusers import FlowMatchEulerDiscreteScheduler
|
| 64 |
+
from typing import Union, Tuple, Optional
|
| 65 |
+
|
| 66 |
+
class FlowMatchEulerX0Scheduler(FlowMatchEulerDiscreteScheduler):
|
| 67 |
+
def step(
|
| 68 |
+
self,
|
| 69 |
+
model_output: torch.FloatTensor,
|
| 70 |
+
timestep: Union[float, torch.FloatTensor],
|
| 71 |
+
sample: torch.FloatTensor,
|
| 72 |
+
generator: Optional[torch.Generator] = None,
|
| 73 |
+
return_dict: bool = True,
|
| 74 |
+
) -> Union[FlowMatchEulerDiscreteSchedulerOutput, Tuple]:
|
| 75 |
+
|
| 76 |
+
if self.step_index is None:
|
| 77 |
+
self._init_step_index(timestep)
|
| 78 |
+
|
| 79 |
+
sample = sample.to(torch.float32) # Ensure precision
|
| 80 |
+
|
| 81 |
+
sigma = self.sigmas[self.step_index]
|
| 82 |
+
sigma_next = self.sigmas[self.step_index + 1]
|
| 83 |
+
|
| 84 |
+
# 1. Compute x0 from model output (assuming model predicts noise)
|
| 85 |
+
x0 = sample - sigma * model_output
|
| 86 |
+
|
| 87 |
+
# 2. Add noise to x0 to get the sample for the next step
|
| 88 |
+
noise = torch.randn_like(sample)
|
| 89 |
+
prev_sample = (1 - sigma_next) * x0 + sigma_next * noise
|
| 90 |
+
|
| 91 |
+
prev_sample = prev_sample.to(model_output.dtype) # Convert back to original dtype
|
| 92 |
+
self._step_index += 1 # Move to next step
|
| 93 |
+
|
| 94 |
+
if not return_dict:
|
| 95 |
+
return (prev_sample,)
|
| 96 |
+
|
| 97 |
+
return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample)
|
| 98 |
+
|
| 99 |
+
pipe = FluxPipeline.from_pretrained("Path/to/SenseFlow-FLUX", torch_dtype=torch.bfloat16).to("cuda")
|
| 100 |
+
pipe.scheduler = FlowMatchEulerX0Scheduler.from_config(pipe.scheduler.config)
|
| 101 |
+
|
| 102 |
+
prompt="A cat sleeping on a windowsill with white curtains fluttering in the breeze"
|
| 103 |
+
|
| 104 |
+
images = pipe(
|
| 105 |
+
prompt,
|
| 106 |
+
height=1024,
|
| 107 |
+
width=1024,
|
| 108 |
+
num_inference_steps=4,
|
| 109 |
+
max_sequence_length=512,
|
| 110 |
+
).images[0]
|
| 111 |
+
|
| 112 |
+
images.save("output.png")
|
| 113 |
+
```
|
| 114 |
+
|
| 115 |
+
## DanceGRPO-SenseFlow (supports 4–8-step generation)
|
| 116 |
+
|
| 117 |
+
comming soon!
|
| 118 |
+
|
| 119 |
+
## Citation
|
| 120 |
+
|
| 121 |
+
If you find this work useful, please cite:
|
| 122 |
+
|
| 123 |
+
```bibtex
|
| 124 |
+
@article{ge2025senseflow,
|
| 125 |
+
title={SenseFlow: Scaling Distribution Matching for Flow-based Text-to-Image Distillation},
|
| 126 |
+
author={Ge, Xingtong and Zhang, Xin and Xu, Tongda and Zhang, Yi and Zhang, Xinjie and Wang, Yan and Zhang, Jun},
|
| 127 |
+
journal={arXiv preprint arXiv:2506.00523},
|
| 128 |
+
year={2025}
|
| 129 |
+
}
|
| 130 |
+
```
|