Meet the child of Project Chimera and Krea 2 Center.
🧠 Brain: Project Chimera (txtfusion layers)
💪 Body: Based on Krea 2 Center's DiT backbone, with additional refinements.
Chimera provides the model's prompt understanding and interpretation, while the image generation backbone builds upon Krea 2 Center. The result is a model that inherits traits from both parents while developing its own personality.
⚠️ Requirements
Please make sure you're using ComfyUI v0.31.0 or newer. Older builds may not include the required W4A8 support.
Low-VRAM users: I also made a W4A8 quantized Qwen 3 VL 4B text encoder, only 2.86 GB. You can find it here: https://civarchive.com/models/2843533/qwen-3-vl-4b-w4a8?modelVersionId=3210230
Description
FAQ
Comments (29)
[ERROR] !!! Exception during processing !!! 'asym_w4a8_int8' .current version of ComfyUi
@pikkby476 yes like I said you need the very latest comfy version…
You will have to git pull or wait till whatever version you use gets the update.
Another great model. I like your work, Bro.
thanks I made release an alternative version tomorrow need to test it a bit more, its the same in a way and not because I changed the txtfusion layers
Do Krea2 loras work on this bad boy?
I did not try because w4a8 has currently an issue with loras its gets incredibly slow.
They do work - though I see no speed difference from an int8 Chimera run on my rig.
@xFennec777 I did not try yet but w4a8 minimax freaked when I did
@TiwazM stacked a character and style and it works
Nice, I'm going to give this a go! Thanks TiwazM
I am all for new tech and making this faster and easier on low end (8gb 3070ti mobile + 32gb ram here). For a baseline: My int8 speeds on Krea 2 are 35s cold and 29s cached. Int4 are 26s cold and 19s cached.
I am getting int4 level output of Character LoRA (uncanny valley) at int8 speeds on my rig for this model. I know it's experimental, so happy to keep testing!
More specifically, quick findings:
- cold generation 34s no Lora, 42s with LoRA
- cached generation 24s no LoRA, 30s with LoRA
- Output respects Character LoRA about the same as an int4, some of the hair and cloth can get a bit patterned
I updated to Comfy 0.31.0 and it supports w4a8
More findings - quality can bounce around. If this gets deeper support in Comfy - could make us GPU poor folks have reasonable generation times!
@xFennec777 yeah I am not sure its a bit heavy merged will try to make an w4a8 of Chimera 1.2 to see if that has the same issues. Thanks for the feedback.
with the Nix Comic lora ?
also seems the soon o be released v1 alternative might work better without lora but only did a quick test.
@TiwazM sweet! I may have time to test tonight or tomorrow, else I can do some deeper testing on Monday :)
Posted a clean run and a Nix Comic on the 1.0 earlier. Pretty solid.
@xFennec777 I did test it a bit with your 1.1 seems not bad but you know how its supposed to look :)
@TiwazM The illustrations are strong with your Center x Chimera, Nix Comic is quite close. I tested a few character LoRA: It would get a likeness, but one that looked off (similar face and body shape, missing freckles, etc - similar to how int4 misses the finer details).
I think the biggest thing with my tests were the slower int8 speeds for int4 handling (int4 isn’t terrible, but just softer)
My findings could be that, though theoretically the model fits my 8gb card, I still get partial loading / model swapping still. Wonder how it compares on a 10gb+ vs an int8 :)
Edit: us 8gb VRAM folks seem to need int8. The text encoders and other overhead push the theoretical bas VRAM to 12gb to ensure full model load and not bump into offloading slowdowns.
Will still use where I can - Center and Chimera are my top 2 models!
Any idea why the int8 and the w4a8 is very slow when is use it in the 3rd loop for upscale to 2k? Every other Krea model I used was more than twice that fast? (16GB VRAM, 32 GB RAM)
@Pressydent no Idea, I render at 2 MPx. Is your workflow available ? Is it the one in the pic you posted ?
@Pressydent are the other krea2 models int8 or w4a8 too ?
@TiwazM I postet a pic. I already used int8 yes. But this was the first w4a8 I used. It takes around 2 minutes for each of the last 4 steps using the w4a8 and 4-5 minutes for each when I use the int8. Normally it takes 20 -40 seconds per step for fp8 or fp16 with the same workflow.
@Pressydent Right, I also tried Krea2-SAT-DirtyRealism in FP8, and on my 16 GB card it was actually slower — roughly 35% more time per step than my INT8.
I didn't really benchmark pass 3, though, because at that point the workflow starts spilling heavily into shared memory on a 16 GB card, which slows things down a lot. So I don't think pass 3 is a very useful model-speed comparison on this VRAM size.
@Pressydent I recreated roughly the same image with my workflow and used the INT8 model.
First pass 100%|██████████| 10/10 [00:10<00:00, 1.02s/it]
2nd pass 100%|██████████| 4/4 [00:18<00:00, 4.71s/it]
3rd pass 100%|██████████| 4/4 [02:13<00:00, 33.43s/it]
Did run it with w4a8 pretty much the same 4 sec difference in total time
feel free to try that workflow
@Pressydent right I uploaded one last one can do the full 3 pass in 4 min optimized the upscaler choice, have a look if you are interested
@TiwazM ty. I will take a look.
It keeps throwing an error. Is there a proper, working process?
you need ComfyUI v0.31.0