The rest of the set (MOSS, Z-Image, img2img, upscaling): tinytensor.gumroad.com
LICENCE — IMPORTANT
Qwen-Image 2.1 is released under the Qwen Research License Agreement: research and evaluation use only, not for commercial output. The licence says "FOR NON-COMMERCIAL PURPOSES ONLY", and commercial use requires a separate licence from Tongyi ([email protected]).
If you need images you can sell, use Z-Image Turbo or MOSS-Image instead — both are Apache 2.0. The other workflows in my 8GB pack cover both of them.
Qwen-Image 2.1 in 4 steps, running on an 8GB laptop GPU at 1024x1024 — about 20 seconds per image.
Everything in this workflow is tuned for 8GB of VRAM, not 24GB.
WHAT YOU NEED
ComfyUI (current build) + ComfyUI-GGUF
qwen-image-2.1-UC-Q4_K_M.gguf (4.6GB) → models/diffusion_models
qwen3vl_8b_int8_convrot.safetensors (9.35GB) → models/text_encoders
qwen_image_2.1_vae_bf16.safetensors (0.68GB) → models/vae
qwen-image-2.1-lightning-4step.safetensors (0.07GB) → models/loras
32GB system RAM
SETTINGS THAT MATTER
steps 4 · CFG 1.0 · euler · simple · denoise 1.0
At CFG 1.0 the negative prompt has no effect — that is normal with this LoRA, not a bug.
If you get noise instead of an image, the LoRA is not connected.
WHY IT FITS IN 8GB
That 9.35GB text encoder cannot sit on an 8GB card. ComfyUI loads it, runs it, then offloads it to system RAM — which only works if the encoder is loaded separately with CLIPLoader instead of inside a combined checkpoint. That, plus the Q4 GGUF weights and the 4-step distillation, is the whole trick.
Measured on RTX 3070 Ti Laptop 8GB / Ryzen 7 6800H / 32GB RAM / Windows 11: 32s on the first run (loads the model), then 20.1s and 20.3s.
The other nine workflows (Z-Image, MOSS 2.5D Anime, IPAdapter character lock, instruction-based editing, 4x upscale) are in a paid pack on Gumroad:
https://tinytensor.gumroad.com/l/comfyui-8gb-vram
Ask in the comments if something breaks — I read them.