BennyDaBall/Qwen-Image-2.1-NVFP4

🚀 Qwen Image 2.1 — NVFP4 for ComfyUI

This is an independent NVFP4 quantization of Qwen Image 2.1. The Qwen team created the original model. The original model and architecture are Qwen's work. I converted the official Comfy-Org/Qwen-Image-2.1 BF16 checkpoint package and packaged the quantized weights, workflows, and validation here. This is not a new base model or an official Qwen release.

The image model AND the text encoder. 12.42 GB for the complete package.

I converted Qwen Image 2.1's image transformer and Qwen3-VL 8B language encoder directly from the official BF16 weights into native ComfyUI NVFP4. The vision tower, embeddings, output head, critical image-model tensors, and VAE stay BF16. Real Blackwell FP4 kernels in both denoising and text conditioning, with an included small runtime patch for the latter. No new nodes.

Built with Qwen. Research/evaluation use only under the included license.

RTX 5090: 6.746 s median for a complete 1024×1024, 40-step run versus 15.629 s with BF16 and 7.553 s with official INT8. Three warm runs per variant, every node executed, including text encoding. This one-prompt benchmark used the encoder patch; full method and limits.

📦 Download once, wire nothing

Tested with ComfyUI 0.36.0, PyTorch 2.14.0+cu130, comfy-kitchen 0.2.35, and an RTX 5090. Use a ComfyUI build with TextEncodeQwenImage21 and native NVFP4 support. Older builds without the Qwen Image 2.1 integration will not work.

File	Where it goes	Size
qwen_image_2.1_nvfp4.safetensors	ComfyUI/models/diffusion_models/	4.20 GB
qwen3vl_8b_nvfp4.safetensors	ComfyUI/models/text_encoders/	7.55 GB
qwen_image_2.1_vae_bf16.safetensors	ComfyUI/models/vae/	0.68 GB

The original BF16 set is 32.44 GB. This package is 61.7% smaller on disk. File sizes are decimal GB, not a VRAM requirement. Keep your existing identical VAE if you already have it.

1. Put the three files in the folders above and refresh ComfyUI's model list.
2. For accelerated NVFP4 text conditioning, apply the runtime patch and restart. Without it, the same models and nodes work, but text conditioning uses dequantized weights and FP32 activations. Denoising still uses NVFP4.
3. Import one of the workflows below using Ctrl+O, then click Run.
4. For editing, use Upload in Load Image to select the included reference image, or your own image.

Working workflow	Purpose
01 — Text to image	1024×1024 product example; change the prompt and latent size
02 — Image editing	One reference, coherent material/color/object changes
03 — Transparent RGBA	Native alpha output; no background-removal node
04 — 2K typography	2048×2048 poster with exact requested lettering

The workflows use core UNETLoader, CLIPLoader (type qwen_image), TextEncodeQwenImage21, KSampler, and VAE nodes. Leave diffusion weight dtype at default. Defaults are 40 steps, Euler, simple scheduler, CFG 1, denoise 1. The encoder node's resolution controls reference resizing; text-to-image output size comes from Empty Latent Image. API equivalents are in workflows/api.

👀 Same prompts, same seeds

All 24 matched pairs are in COMPARISONS.md. Download gallery.html and its assets folder together for the offline comparison slider. See PROMPTING.md for text, editing, lettering and transparency tips.

See VALIDATION.md for the test set, measured results, precision tradeoffs, and execution requirements. The comparisons use BF16 source weights against this complete accelerated NVFP4 package. Quantization changes results; this is not a promise of pixel-identical output or a universal quality ranking.

🧠 What's actually quantized?

Component	Policy
Image transformer	192 attention Q/K/V/output and fused MLP gate/up/output matrices: NVFP4
Qwen3-VL language backbone	252 attention and MLP matrices: NVFP4
Vision tower, token embeddings, language head	Original BF16 tensors retained
Image-model input/output, timestep, modulation, normalization	Original BF16 tensors retained
VAE	Original BF16 file, unchanged
Scales	E4M3 FP8 blocks of 16, FP32 tensor scales

Native packed E2M1 FP4, native scale layout, and per-layer comfy_quant metadata. This is a mixed-precision NVFP4 stack, not every parameter forced into four bits. No GGUF loader, no requantization of the INT8 release, no dropped vision tower. No calibration or fine-tuning was used for this conversion.

🔧 Reproduce it

Source repository: Comfy-Org/Qwen-Image-2.1, revision ace0edeb3791a594ddfa36ed5f41a178a394e921.

Use the Python environment of the tested ComfyUI installation:

python tools/convert.py path/to/qwen_image_2.1_bf16.safetensors diffusion_models/qwen_image_2.1_nvfp4.safetensors --component dit --comfy-root path/to/ComfyUI
python tools/convert.py path/to/qwen3vl_8b_bf16.safetensors text_encoders/qwen3vl_8b_nvfp4.safetensors --component encoder --comfy-root path/to/ComfyUI

The converter uses installed Comfy Kitchen operations, validates matrix alignment, and emits tensor manifests with source/output hashes and reconstruction error. tools/verify.py SOURCE CONVERTED checks the stored format, checkpoint hash, and exact equality of every protected tensor. Final file checksums are in SHA256SUMS.

⚖️ License & credits

Qwen created the base model; Comfy-Org provided the native BF16 checkpoint package; ComfyUI and Comfy Kitchen provide loading and GPU execution. Independent conversion, workflow packaging, and RTX 5090 validation by BennyDaBall_OG.

The Qwen Research License limits use to non-commercial research and evaluation; commercial use requires a separate license from Qwen. Read the full license and NOTICE. The included ComfyUI patch remains GPL-3.0. This repo is not affiliated with Qwen or Comfy-Org.