FP8Enjoyer/FLUX.3-Klein-Step-Distilled-MXFP8

📌 Note — Downloading the Model

Please clone the complete repository rather than downloading individual files. This ensures you receive the model checkpoint, configuration, license, and required files together.

git clone https://huggingface.co/FP8Enjoyer/FLUX.3-Klein-Step-Distilled-MXFP8

Prerequisites: Git, Git LFS, and Git Xet for downloading large model files.

FLUX.3 Image Klein SD — MXFP8 (ComfyUI)

Unofficial community quantization of the distilled 16-step FLUX 3 Image Klein image transformer by Black Forest Labs. Not affiliated with or endorsed by Black Forest Labs or Comfy-Org.

This release contains only the diffusion transformer. Obtain the corresponding Qwen3-VL text encoder and FLUX VAE separately from the official ComfyUI repository.

What changed

• Original: flux3_image_klein_sd_bf16.safetensors (~11.1 GB).
• Output: flux3_image_klein_sd_MXFP8.safetensors (5.7222 GB).
• 273 two-dimensional weight matrices converted to MXFP8 E4M3 with E8M0 scales per 32 elements and blocked ComfyUI scale storage.
• 86 normalization scale tensors preserved as BF16.
• 5,548,670,976 of 5,548,681,984 parameters quantized (99.999802%).
• File holds 905 tensors: 273 FP8 weights + 273 scale tensors + 273 ComfyUI markers + 86 BF16 norms.
• The original tensor names are retained; no architecture changes or training were performed.

The checkpoint passed a structural audit (tensor names, shapes, types, quantization markers, scale layouts) and successfully generated nine evaluation images with ComfyUI on an RTX 5070 Ti Laptop GPU (Blackwell SM120) and ComfyKitchen 0.2.37, after applying the compatibility patch below. Local performance measurements and matched BF16/MXFP8 image comparisons are documented below; quantitative perceptual-quality validation remains to be done.

Local performance benchmark

Measured on an NVIDIA GeForce RTX 5070 Ti Laptop GPU (12,199 MB reported VRAM) with 32,125 MB system RAM, Windows, ComfyUI 0.41.0, PyTorch 2.10.0+cu130, and ComfyKitchen 0.2.37. All sessions used LOW_VRAM mode, DynamicVRAM, async weight offloading with two streams, and PyTorch attention.

Each configuration completed the same nine-prompt sequence in a separate ComfyUI session. The first execution is reported separately; warm results cover the remaining eight executions. Warm executions use different prompts, rather than eight repeats of one prompt.

Matched generation settings

The original PNGs retain their ComfyUI prompt and workflow metadata. For all nine BF16/MXFP8 pairs, comparison of the embedded execution graphs found only one changed input: the diffusion model filename. Both used weight_dtype = default.

Setting	Value in both BF16 and MXFP8 outputs
Output size	1440 x 1440, 1:1
Resolution selector	2.0 MP target, multiple of 8
Seed	1
Steps	16
CFG	1.0
Sampler / scheduler	euler / simple
Denoise	1.0
Batch size	1
Negative prompt	Empty
Text encoder	qwen3vl_4b_fp8_scaled.safetensors, loader type flux3, device default
VAE	flux2-vae.safetensors
Positive prompts	Identical within each pair; nine different subjects

The forced-FP8 baseline loaded the original BF16 checkpoint with weight_dtype = fp8_e4m3fn. Its timings and weight dtype are recorded in the console log; those earlier output PNGs were deleted and are not included in the visual evaluation or PNG metadata audit.

Timing and reported weight footprint

Metric	Original checkpoint, forced FP8	Original checkpoint, BF16	MXFP8 checkpoint
First execution, including loading	46.77 s	52.99 s	25.76 s
Warm execution, mean (n=8)	38.74 s	38.89 s	23.70 s
Warm execution, range	36.16-41.67 s	38.15-40.14 s	22.38-26.43 s
Warm throughput, derived from mean execution time	1.55 images/min	1.54 images/min	2.53 images/min
Warm sampling, mean logged seconds/step	2.133 s/step	2.155 s/step	1.275 s/step
Diffusion weights reported as staged	5,291 MB	10,583 MB	5,456 MB

On this setup, MXFP8 achieved 1.64x the warm throughput of BF16 (39.0% shorter mean execution time). The first execution was 51.4% shorter, and the reported staged diffusion-weight footprint was 48.4% smaller.

Measurement boundaries: First-execution time includes model loading, text encoding, sampling, VAE decoding, and saving; standalone model-load time was not measured. ComfyUI application startup is excluded. Throughput is derived from ComfyUI execution times and excludes queue waiting and user interaction. The staged-weight figure is a loader-reported footprint, not peak VRAM or system RAM use; peak runtime memory was not recorded. These are single-session observations on one laptop, without repeated sessions, controlled clocks/temperature, or statistical significance testing. Startup background tasks were still completing during some executions, so the results should be treated as a local comparison rather than a general hardware speed claim.

The per-execution timings and paired PNG metadata audit, full prompts, and original-file SHA-256 hashes are included for inspection.

Visual evaluation: BF16 vs MXFP8

All nine matched pairs are shown below: BF16 on the left, MXFP8 on the right. At normal viewing size, and in selected original-resolution detail crops, the outputs appear similar in overall color, contrast, and resolved detail. No consistent washed-out appearance or broad detail loss was apparent in this small set. The largest content change is in the fruit arrangement; wheel details, faces, hair, signs, and other small features also vary.

This is a qualitative visual inspection, not a blinded study or a quantitative perceptual benchmark. The outputs are not pixel-identical, and these nine prompts at one seed do not establish quality parity across other prompts, seeds, resolutions, or hardware.

Portrait, dog, and Tokyo street

Cafe, mountains, and fruit

Car, butterfly, and anime

The comparison sheets use equally resized previews with JPEG compression and no color or contrast adjustments. Use the original PNG links below to inspect full-resolution output and embedded metadata. The detail-crop sheet is a lossless PNG containing selected original pixels without resizing.

Pair observations and original PNGs

Subject	BF16 original	MXFP8 original	Visual observation
Sunflower portrait	PNG 00001	PNG 00010	Small facial and hair changes; skin texture and warm lighting remain visible.
Golden retriever	PNG 00002	PNG 00011	Very similar pose, fur detail, grass, and color.
Tokyo street	PNG 00003	PNG 00012	Signs and reflections vary; neon intensity and dark shadows remain comparable.
Cafe interior	PNG 00004	PNG 00013	Small differences in cup geometry and foam pattern; similar lighting and surfaces.
Mountain landscape	PNG 00005	PNG 00014	Similar mountain composition, snow detail, trees, and lighting.
Tropical fruit	PNG 00006	PNG 00015	Largest composition change: fruit positions and papaya shapes differ; both retain detailed surfaces.
Sports car	PNG 00007	PNG 00016	Wheel design and body details vary; reflections and sunset remain comparable.
Butterfly	PNG 00008	PNG 00017	Wing markings differ; fine wing texture remains resolved in both.
Anime portrait	PNG 00009	PNG 00018	Small changes to face, hair, and ornament details; crisp linework and shading remain visible.

Installation (Windows / ComfyUI)

1. Place flux3_image_klein_sd_MXFP8.safetensors in ComfyUI/models/diffusion_models/ (or a configured extra model path).
2. Ensure your ComfyUI Python environment has working Blackwell MXFP8 support in ComfyKitchen. Version 0.2.37 was used in the successful local run; this is not a guarantee for every setup.
3. Apply the compatibility patch (see below) with ComfyUI closed.
4. Start ComfyUI and choose flux3_image_klein_sd_MXFP8.safetensors in the Load Diffusion Model / UNETLoader node. Leave weight_dtype = default.
5. Use the official FLUX 3 Klein distilled workflow and matching text encoder + VAE. The distilled model normally uses 16 steps, with no separate CFG guidance requirement.

One-command patcher

patch_comfyui_mxfp8.py uses only Python's standard library. Run:

py -3.13 patch_comfyui_mxfp8.py

When prompted, enter the ComfyUI installation root, e.g. C:\path\to\ComfyUI (not models or venv_adv). The script locates comfy/quant_ops.py, makes a one-time quant_ops.py.mxfp8_backup, verifies syntax, and updates only this MXFP8 activation-quantization call:

# before
ck.quantize_mxfp8(tensor, pad_32x=needs_padding)
# after
ck.quantize_mxfp8(tensor.contiguous(), pad_32x=needs_padding)

Why? The stock FLUX 3 forward path can supply a non-contiguous activation after a transpose. ComfyKitchen's MXFP8 CUDA kernel requires contiguous input. This patch is runtime-only; it does not change the checkpoint or produce new weights. It may become unnecessary after an upstream update.

Additional commands:

py -3.13 patch_comfyui_mxfp8.py --check
py -3.13 patch_comfyui_mxfp8.py --restore

The script is idempotent, refuses unknown ComfyUI source layouts, and does not overwrite a different backup. To use it without a prompt, pass --comfyui "C:\path\to\ComfyUI". Restart ComfyUI after patching or restoring. If the patch reports an unsupported source layout, do not force it.

Compatibility, limitations, and reproducibility

• This is a ComfyUI-native MXFP8 Safetensors export, not a GGUF or Diffusers checkpoint. config.json records release metadata; the ComfyUI loader uses the Safetensors embedded metadata.
• The official Comfy Quants project provided the MXFP8 conversion function; weights were converted from the official BF16 distilled checkpoint.
• MXFP8 activation quantization and GPU execution depend on ComfyUI, ComfyKitchen, PyTorch, GPU architecture, and driver compatibility.
• The patch is tailored to the documented comfy/quant_ops.py implementation and could need adjustment on later ComfyUI versions.
• Performance measurements and nine matched visual comparisons are reported for one configuration. Peak runtime memory, repeated-session performance, and quantitative perceptual-quality validation against BF16 remain unmeasured.

Licensing and attribution

Original model: © Black Forest Labs, FLUX 3 Image Klein, licensed under the FLUX Kommunity License v1.0. The checkpoint is a quantized derivative, and the license's restrictions and distribution conditions continue to apply. Read the full original license and official NOTICE.

Modification notice: Original 16-step distilled BF16 diffusion-transformer weights were converted offline to MXFP8 E4M3 with E8M0 block-32 scaling, maintaining 86 normalization tensors in BF16. This is not an official BFL checkpoint.

ComfyKitchen is separately licensed under Apache-2.0. This patch script does not bundle ComfyKitchen source or code from Comfy Quants.