ComfyUI Simple Flux - Turbo UnSampler & Inpainting
Added 2024-10-17 12:11:51 +0000 UTCVideo - https://youtu.be/-qVDWE0A1ek
Loads of changes in this one.
With the recent release of the AliMama beta inpainting controlnet, I figured I'd try it out, and now you can too!
The new Input Image section let's you easily pick from a variety of image-related options:
Enable IMG2IMG will use the image latent instead of an empty one
IMG Megapixel Scale sets the image size in megapixels
IMG Noise has a limited range of around 0.8 to 0.95, with higher numbers changing the image more
Enable Inpainting will require you to add a mask (via mask editor) to your image to represent the area you want to inpaint.
Composite Inpaint will add the inpaint result back over the original image. Sometimes this is good, sometimes bad. Inpainting will always change the entire image (not just the masked area) so fine details (such as small text) always get changed. The composite counteracts that... however it can also make the masked area appear much more obvious :|
The Rectified Flow Unsampler will first turn the image into noise and then back into an image again. This allows for small or large changes... or even really strange things!
Example Screenshot of a completed workflow run:
AliMama
The defaults for this uses 2 recently released models:
https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha
https://huggingface.co/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta
The "Turbo Alpha" is much the same as Hyper-SD, but the filesize is much smaller. Results are obviously different to Hyper, but as to if they are better/same/worse...? I guess I'd probably go with about the same, though the smaller file size may tip it slightly in Turbo's favour.
The Inpainting Beta is better than the alpha, but I'm still getting plenty of obviously masked areas in many cases. Some things work really well, such as changing faces, but where the shape doesn't match so well your mileage will vary.
Give it a go and let me know what you think!
Rectified Flow Unsampler
https://github.com/logtd/ComfyUI-Fluxtapoz
A WIP but still really good fun :)
Other Notes
1. As with many models, they use generic filenames - diffusion_pytorch_model.safetensors - for everything. When downloading the files, it's a good idea to change the filenames to reflect what the model is. That way you know for later! If you want to use the same names as me, I used:
Alimama-Ctrl-Inpainting-Beta.safetensors
Alimama-Flux-Turbo-Alpha-8steps.safetensors
As usual, controlnets go into "ComfyUI/models/controlnet..." and loras into "ComfyUI/models/loras..."
2. Switches are fun ;) You can still deliberately break things by changing the value of the muted "composite" node when you're not using inpaint mode. Not sure how I can get around that yet!
3. I still almost kinda prefer IMG2IMG over inpainting, though I think that could change with an even better inpainting model!
4. The UU workflow also includes Ultimate Upscale
Comments
Dapper Node Rodent Suggests Alimama, (comma) Llama To Avoid Subscription Conniptions, Drama It's hard to keep up with the Turbo variants and what not, but if my Stable Diff-era self heard me say that, he'd kick me in the nards and say something like "well Dall-E is right over there, loser". And he's right. We have to fight for open source, and it has absolutely nothing to do with naked ladies. (walks away with 4TB sack full of loras, checkpoints, and duplicates of both - slightly renamed - Joan Rivers Real XL falls out) Oh that one isn't mine, that's a joke. Can you toss that in the Recycle Bin or whatever Linux does with trash?
NetDude
2024-10-18 06:16:49 +0000 UTC
