Do you want to get the Flux.2 working with 12 Gb VRAM?
Then this workflow is just what you need!
downloads:
https://huggingface.co/Comfy-Org/flux2-dev/blob/main/split_files/vae/flux2-vae.safetensors
https://huggingface.co/Comfy-Org/flux2-dev/blob/main/split_files/text_encoders/mistral_3_small_flux2_fp8.safetensors
https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
Update: if you switch off the eject_models then set VAE to CPU
Description
yes, this WF is working on low VRAM!
FAQ
Comments (22)
I was sure low vram means 8GB....... so it's a no-go for me
well, the method used in this WF (maybe) will be working on 8 Gb VRAM if you have enough RAM to offload to - I simply was not able to test it on 8 Gb - hence the number is: 12 Gb (tested)
small FLUX2 will be out later ;)
@sevenof9247 then this WF will be usable for even smaller VRAM in that case (if only ...)
We have arrived at the time when 32gb is the low ram option.
When I firs read the "mistral ... small" filename, I was not expecting the fking 18GB ahh kiddo 💀
text encoder model gets out of the memory before the sampling starts
I worked on it myself, RTX 3060 Ti 8GB VRAM, 32GB RAM, I used Flux 2 Dev Q2, I will test Q4, because as I said it worked but it took a while to generate the image.
Where to get clownsharksampler_beta? GitHub master branch offers without beta.
@0l1v1aR0551 did a fresh pull but Beta Sampler is missing
@zoom83 you need beta57 - not sampler -> scheduler
Invalid Workflow against zod schema: Validation error: Invalid format.Must be " github-user-repo-name" at " nodes[0].properties.aux_id"
the error is obviously on your side (something with Comfy)
nah every other workflow I am testing is working just fine, only this one in particular gives this error.
@blhll this WF uses some nasty nodes for offloading - they are (probably) to blame
Not job!
I tested this workflow with flux2‑dev‑Q4_K_S.gguf on an Intel Arc A730M (12 GB VRAM) and made sure to load all the required files (model, VAE, CLIP, etc.). However, I still keep getting the error “too many values to unpack (expected 4)” when running the CLIPTextEncode node. It looks like an incompatibility between Flux2 GGUF and the current MultiGPU patch. Has anyone found a reliable fix or workaround to make this workflow run correctly on Arc GPUs?
there are other gguf loaders (nodes = at least from 3 different authors) - you can try them too
I have had problems with MultiGPU/Distorch with ComfyUI 0.3.76 onwards. Try going back to 0.3.75. It looks like Linux users with low VRAM are going to be stuck on that version of a while.
You can get previews working if you disable animated previews in the settings. However it's still slow garbage and Q2_K gives far inferior results to other prevailing models that fit well at Q4_K_M or even the photorealistic Illustrious models that fit the whole fp8 model. Just wait for Klein instead if you want Flux.2.
What's the generation time on nvidia 16gb or 12gb 40 series gpu with this workflow? I have a 16gb 4060ti and 64gb of ram
when importing workflow, the warning "Invalid workflow against zod schema" appears.: Validation error: Invalid format. Must be 'github-user/repo-name' at "nodes[0].properties.aux_id""