This is a full checkpoint with 8-bit precision (FP8) based on Flux Fusion (which gives best results for me, and also there are variations with some other models). I made the checkpoint just for my own convenience. Unet, Clips and VAE merged to the checkpoint are described in this article.
It's okay to have 4GB VRAM and 32GB RAM to use the checkpoint. Simple testing ComfyUI workflow (.json file) is attached here as "Training Data" in .zip archive. 1024x1440 image generation takes ~80 seconds on my notebook with 4GB 3050 and i5-11400.
8 steps are enough to generate hi-res sharp images.
Just download and run it.
Enjoy.
Description
Based on FluxUnchained
NSFW friendly
FAQ
Comments (27)
When I load the example workflow, it says the "CheckpointLoaderSimpleMultiGPU" node is missing. Is it from a custom extension to ComfyUI?
And can I use the standard Load Checkpoint instead? (I don't have multiple GPUs)
For some reason standard checkpoint loader leads ComfyUI to crash on my notebook with 4Gb VRAM, and that custom loader do the job well.
You should install ComfyUI manager if not yet.
https://civitai.com/models/71980/comfyui-manager
Then select in manager to install missing nodes, it will do it for you.
@mistporyvaev For some reason loading missing nodes didn't work for me, but I found the extension with that node. Also I use an Intel GPU (A770), which complicates things. I managed to get the multi-GPU loader node to work by making some changes in the Python code (normally it only accepts CUDA or CPU), but I haven't noticed any difference in performance compared to the standard loader node.
And the performance is not great, it takes almost all the VRAM (16 GB), and the generation is pretty slow. It varies though, usually it's over 30 s per iteration (512x1024), but sometimes it's much less, like 10 or 20 s, and once I even got something like 3 s. So far I haven't managed to get any Flux-based model to work consistently with reasonable speed, that's probably because I use Intel instead of NVIDIA like normal people. SDXL-based models work very well though (less than one second per iteration on an 1024x1536 image).
@PawelS Good news though, now I know that Flux is runnable on Intel Arc videochip ✌️
Update: I found a model that works fast for me: https://civitai.com/models/920261?modelVersionId=1030160 (the Q3 version), a 1280x720 image takes about 20 seconds to generate.
@PawelS 3 bits vs 8 bits, Flux.1S vs Flux.1D. What about quality and details?
@mistporyvaev It's good enough for me.
@PawelS could you show something generated by Q3 model?
@mistporyvaev https://civitai.com/posts/11607121
@PawelS Nice pictures. Thank you
I don't quite understand how you generate images in 80 seconds. I have a laptop with an AMD Ryzen 7840HS CPU, RTX 4060 GPU with 8GB VRAM, and 64GB RAM installed, and It took me 377 seconds to generate an image with the same settings that came with your workflow. So it takes me almost 5 times longer with twice the VRAM.
EDIT: 287 seconds on second run and 191 seconds on third run. Weird.
EDIT: I changed sampler and scheduler to euler/beta and resolution to 768x1024, and that takes only about 45 seconds @90% VRAM usage and 41% RAM usage, and the quality is just as good. Then I use Topaz Photo AI to upscale the image.
I use ComfyUI last weeks. That's all, there are no other secrets or tricks. It just works =)
First run always takes a longer time because of models loading.
1024x1440 generation by 8 steps lasts ~80secs on RTX3050 with 4Gb VRAM and less than 40secs on RTX3070 with 8Gb VRAM.
You should update drivers and OS maybe.
And you could try this workflow
https://civitai.com/articles/10486/low-vram-flux-comfy-workflow
@mistporyvaev nah, for me on an rtx 3070 8vram gets stuck on reconnecting due to high demand, so doesnt work on every environment bro
@anasexta42122 your VRAM config is more than enough for successful generating via Flux. How much RAM you have? Which software you use?
@mistporyvaev oh sorry for the late reply, 16g RAM and im working on comfyui, but i could never get video gen to work actually nor flux alone as image gen
@anasexta42122 you should try to set enough amount of swap memory on the fastest disk you have. Try to set 64Gb of swap memory.
Very natural imitation of photos from mirrorless cameras. Simply amazing!
Hi! Will this work on Forge UI with the same speed? I only have RTX 3060 ryzen 9.. I am fairly new to this, And I find Forge ui more user friendly than Comfy ui..
Hi! I have mobile RTX 3050 and I know nothing about ForgeUI 🤐
I hope 3060 will be faster than my setup.
It works amazingly well on forgeui. no need for clip, vae and text encoder. just the model is enough.
@dfhd yep. It's AIO = All In One
@dfhd what GPU do you have?
@mistporyvaev till 10 days back gtx 1650 (4GB DDR6), now upgraded to 3060Ti (8 GB DDR6). Thank you very much for this checkpoint. It is amazing. also this has become my regular goto model now. used shuttle mixes before finding this. Also the one based on flux fusion is much better than the one based on flux unchained.
Unfortunately I’m bloody dumb and cannot figure out how to actually run this. The workflow you shared with noxxious is missing 4 nodes for me and ComfyUI Manager for some reason fails to load into the actual UI (I put it in custom_nodes, all other custom things load correctly but not Manager). Halp
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.