No, the original Qwen-Image_bf16 is ~40GB, which will not fit in VRAM of a 5090. The fp8 version will (~20GB), and produces great results locally. If you need to go smaller, there are a number of GGUF Quants now too.

theally Oh, ok. I see there's a Q8 quant. Is that better quality than FP8?

Also, I'm getting halftone pattern texture all over my images. It's not always noticeable, but since Qwen produces quite soft images I'm sometimes using image sharpening nodes in ComfyUI and this really makes this pattern visible. Do you know how I can get rid of this pattern?

theally On a 5070 ti, with 16gb vram, are there gguf options?

Thank you so much for making this available to a wider audience :)

theally Ok, so I've tried the BF16 model and it did work on my setup! Not sure how, maybe Comfy offloads parts of the model to system RAM on the fly. I didn't notice major slowdowns in generation time. Maybe some slight delay before it initiates inference.

I am using 4bit gguf K_S version of qwen image on my RTX4060 8GB with 16GB RAM, it take about 330 seconds to generate an image.

We'll be adding Qwen-Image-Edit support soon!

Why not? It seems to be working at the moment.

There is a ahego qwen lora but it looks lower quality. The current MCNL lora doesnt seem to do tongues well either

If you have their workflow and are literally running it but aren't getting the same results they could still be using different loras, different prompt, or different models. They also could have upscaled it, or altered the prompt or the negatives in ways that weren't shared.

I can get perfectly clear photos with no particular positive brute-forcing with 8-steps, res_2s, bong_tangent. Also (eta 0.5, steps_to_run -1, denoise 1.17, cfg 1.14, model shift 1.04, sampler standard). (yes those denoise and cfg are pretty schizo values, but I got this workflow from here: https://civitai.com/models/1851673/mcnl-multi-concept-nsfw-lora-qwen-image)
pretty sure this image specifically: https://civitai.com/images/95469890

Also, a common strategy right now is to create an image with Qwen and then upscale it with Wan2.2. I don't exactly know the photo you are referring to, so I don't know your issue. But I'd say what got my generations to the clarity of Wan2.2 was a combination of the Lightning Lora and realism loras. Otherwise, I don't think any of my phrases or negatives are all that important.

Here's a brute-force negative: "blurry, out of focus, shallow depth of field, low quality, bad quality, low detail, mutated, jpeg artifacts, compression artifacts, bright colors, overexposed, static, blurred details, subtitles, style, artwork, painting, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, malformed limbs, fused fingers, three legs, bad neck anatomy, walking backwards, model, idol, face too pretty, neck completely rotated, head completely backwards,"

что-нибудь поменялось?

@prot1wow37 Нет она выдает ошибку при создании запроса. Красным значок постоянно.

@meikerAI3d можешь пожалуйста скинуть свой workflow? у меня RTX 3050 8gb, но ComfyUI постоянно ошибку "переподключение" выдает. Flux, Chroma без проблем работают, только Qwen нихуя не запускается

Я на Forge запустил

Check the official ComfyUI documentation (https://docs.comfy.org/tutorials/image/qwen/qwen-image). There should be an example workflow that automatically downloads the text encoder, as well as a direct download link for the text encoder (https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors).

@VV24 im sorry how did you found the link of that fp8 text encoder? i was looking for an hour and found only 3split part 1 part 2 part3.. and i was confused. thanks!!!

@gemstonebro No worries, this stuff can be pretty confusing. The download link is right there in the documentation I linked, in the section "Model Download", along with all the other required model files: https://docs.comfy.org/tutorials/image/qwen/qwen-image#2-model-download

Hey, I've raised the issue with the Dev team and am awaiting information. I apologize for the inconvenience, we'll get it back up ASAP!

It hasn't worked even once for me in the past month.

Not yet, no. It's being worked on!

@theally Must be a really big issues if it has take over 4 months to fix when it was working before! ?....

@J1B Yup, difficult issue to solve. We've had to bring in more compute (GPUs), and we're close to relaunching, with the latest model, and Qwen Image Edit support.