These are direct conversions from the original BF16 weights, for use in ComfyUI with the ComfyUI-GGUF custom node by city96.
Since these are quantized models, all original licensing terms and usage restrictions remain in effect.
Why this model
Qwen-Image-Flash generates in 4 steps instead of the usual 20+, with quality close to the base model. The distillation keeps the base transformer architecture and its 20.43B parameters, so it drops straight into existing Qwen-Image workflows — you only change the sampler settings.
Quantizing it brings a 4-step 20B model into range for consumer GPUs.
Files
Quant
Size
Notes
Q8_0
~21.7 GB
Near-lossless, reference tier
Q6_K
~16.9 GB
Excellent quality
Q5_K_M
~14.6 GB
Very good quality
Q4_K_M
~12.3 GB
Recommended — best quality/size balance
Q4_K_S
~11.6 GB
Slightly smaller, slightly softer
Q3_K_M
~10.0 GB
Usable, some detail loss
Q2_K
~7.4 GB
Smallest, noticeable quality loss
Sizes are approximate; check the file list for exact figures.
Required companion files
The GGUF replaces only the diffusion model. You also need the standard Qwen-Image text encoder and VAE — the same ones used by any existing Qwen-Image workflow:
This is a distilled model with a fixed sampling schedule. These settings are not optional — using base Qwen-Image settings will produce poor results.
Setting
Value
Steps
4
CFG
1.0
Sampler
euler
Scheduler
simple
ModelSamplingAuraFlow shift
3.0
Resolution
1024 x 1024
Why CFG 1.0: the teacher used CFG 4.0 during distillation and that guidance is baked into the student weights. Applying CFG again double-applies it. For the same reason, do not add a CFGNorm node — with CFG at 1.0 there is nothing to normalize.
Why shift 3.0: the model ships with a static shift-3 FlowMatch Euler schedule producing sigmas 1.0, 0.9, 0.75, 0.5, 0.0. Shift 3.0 over 4 steps reproduces that trajectory exactly. Higher shift values (13.0 is the common base Qwen-Image default) break the schedule and cause artifacts.
Resolution: 1024 x 1024 is the only resolution NVIDIA tested. Other sizes work if width and height are divisible by 16, but quality varies and higher resolutions may show repeated structure.
Node setup
Unet Loader (GGUF) — select the GGUF
ModelSamplingAuraFlow — shift 3.0
Load CLIP — qwen_2.5_vl_7b_fp8_scaled.safetensors, type qwen_image
Load VAE — qwen_image_vae.safetensors
KSampler — 4 steps, CFG 1.0, euler, simple
Leave the negative prompt empty
Conversion notes
Converted from the original BF16 safetensors with a streaming converter (never loads the full model into RAM), then quantized with llama-quantize.
Precision-sensitive layers are preserved at F16 rather than quantized:
This matters more than usual on a four-step model: the timestep embedding has only four sigma values to distinguish, so quantization error there degrades every step of the schedule. llama-quantize does not preserve these automatically, so they are restored after quantization and verified per file.
Every file was checked after conversion to confirm the sensitive layers are stored at F16 and that the tensor count and architecture metadata are intact.
