👑 Krea2 DMerge V3.0 IS HERE! The Leap to True Mastery! 👑
🚨 WE SPRINTED PAST V2! This iteration is more refined than ever. 🚨
✨ WHY V3 IS YOUR NEXT GO-TO CHECKPOINT ✨
🖼️ FLAWLESS BASE PERFORMANCE: The core weights deliver top-tier image quality even without heavy LoRA reliance.
🛡️ ROCK-SOLID INTEGRATION: Testing expanded further! Experience unbreakable stability when paired with external LoRAs.
⚡ NEXT-LEVEL SPEED (INT8 + BF16): Enhanced the speed pipeline. Now optimized with INT8 CONVROT plus the advanced stability of BF16 loading! Perfect for speed or maximum detail.
Jump in, load this masterpiece, and prepare to be wowed. Showcase your creations below—we want to see what Krea2 DMerge V3 can do! 🌟
See the upgraded visuals in the gallery! 👇
🙏 Massive love and thanks to the legends behind this merge:
This doesn't exist without the brilliant work of GBRX and cielbleu9991 . These are my all-time favorite creators and their work is ALWAYS outstanding. Seriously, check out their originals and support them! 💙
Description
FP8 for those that can't use convrot.
FAQ
Comments (19)
I like this model.
You absolutely nailed it this time, great quality of the bf16 model!
Thank you—your work is very easy to work with.
I've tried the int8 V2 and int8/fp8 V3 and get nothing but black outputs. :(
using this workflow - https://civitai.red/models/2169381/krea2-turbo-workflow-upscale-reference-editing-uncensored
Maybe try the regular template workflow first or the linked one in suggested resources and compare. Your ComfyUI is up to date?
@daceheg192491 I updated yesterday due to getting black outputs with v2 int8, and is why I also tried the v3 fp8 version. I'm generating just fine with several other checkpoints.
just tried a barebones workflow and same issue, even tried the default clip/vae. bizarre... oh well, thanks anyway. :)
Same in webui forge neo with every version of this model; literally every other Krea2 checkpoint works fine :(
@sehrnaise7346865 Same situation.
@daceheg192491 @sehrnaise7346865 @rick3d I think I've figured out the problem.
diffusion models recently converted etc. (whatever the terms are) in ComfyUI no longer seem to be compatible with the Load Checkpoint node... model works fine after using Load Diffusion Model node. presumably forge neo is using what is essentially Load Checkpoint by default.
was having an issue with another model recently uploaded and changing node fixed my issues with both. which was a massive inconvenience, I had to move them to another folder!
@dow185 afaik the difference between the two is whether or not it loads a bundled VAE and CLIP encoder with the model.
@daceheg192491 did you bundle a VAE with your checkpoints by any chance?
@dow185 Are you aware of any fix when using forge neo?
@sehrnaise7346865 No. For ComfyOrg and all other krea2 diffusion models it is usually a stand-alone vae/clip as far as I know.
@sehrnaise7346865 Updating forge neo and installing the Stable-yogi/sd-forge-int8-convrot extension also didn't help :(
@schroeck had read somewhere adding even "lowres" to a negative prompt can help? Euler / Beta for sampler? I only use Comfy so I'm not really sure. V3 was built in the same way using the same tools as V1 and at least in comfy, I'm using the default Load Diffusion Model node w/o issue and w/o a single black output image. Couple things to try:
1. Check the Main UI Mode / Architecture Preset: Ensure the top UI dropdown is explicitly set to Krea2 (or your backend's corresponding MMDiT preset) so Forge Neo correctly maps the layer keys.
2. Assign the Text Encoders and VAE explicitly: :
Text Encoder: https://huggingface.co/Comfy-Org/Qwen3-VL/blob/main/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
3. Verify SageAttention: Beyond file structures, Krea 2 may have a known attention-mask incompatibility with SageAttention in Forge Neo. If SageAttention is active in your command line args or UI setting, disable it or switch to standard sdpa or xformers—SageAttention causes Krea 2 attention calculations to resolve strictly to NaNs (black frames) regardless of how the model is bundled.
@daceheg192491 Looking at the logs, I am actually already getting errors during model loading in webui. Maybe your combination of INT8 + BF16 that you mention is somehow not supported?
@sehrnaise7346865 They are individual. BF16 is created first and then the others from that one. I can see if I can create with that has the vae+clip also if that's the only way forge neo works I guess. Share any log lines that might be relevant please.
@schroeck there's an underlying .py file it's calling in order to load the model, so if you can switch out the file/change the code you could get it to work. I've used Grok for this sort of thing multiple times and it's been pretty good about it.
I expect this will get an official fix eventually, but it might happen on comfyui's end, in which case forge neo would simply not be able to use Models created (using comfyui?) during that time frame (unless there's an official, alternative loader that is equivalent to Load Diffusion Model).
I'm just a novice so I don't even know what I'm talking about - goodluck!
What's the difference between original and this checkpoint?
This is a merge of multiple checkpoints at different weights + loras into one. Hopefully taking the best aspects of all. V3 uses different base checkpoints than V1.