My advice, use:
V2 & V3 for graphic design, illustration, and painting
V1 & V4 for photorealism
V4 (bf16 & Int8 ConvRot)
There's no point in chasing two hares at once, Variant 4 is oriented exclusively towards the photorealism, The pictorial aspect will be considered exclusively in subsequent versions..
V4 has been completely uncensored, if any censorship persists, it stems from the Qwen3VL text encoder, though this remains very rare.
The colors have been kept to a very neutral and natural palette.
V3 (bf16 & INT8 ConvRot) :
Improved skin texture for greater realism
Optimized lighting effects, more subtle and with better contrast
Enhanced painterly visual style
Corrected color shift toward warm tones, color rendering is more neutral and realistic
Feedback on V3
V3 is a lot more restrictive
Yes, the "unlocked" feel has been diluted; I’ll need to revisit that in future versions.
Colors are less realistic.
Yes, that aspect still needs fine-tuning in subsequent versions.
Less realistic than V1
That’s possible, given that I’m aiming to finalize a versatile model, one that is realistic for photography yet also graphic in style.
KREA 2 TURBO Vs KREAMANIA
V2 (bf16 & fp8) : Improvement of version 1 by merging an unpublished personal LoRAs to optimize lighting, textures, and graphic effects.
V1 (fp8) : The removal of certain restrictions compared to the model turbo fp8 scaled, along with improved realism and nudity rendering. The goal wasn't to focus on an NSFW model (even though it is capable of that), but rather to unlock the full range of creative possibilities.
Recommended ComfyUI extensions:
https://github.com/ethanfel/ComfyUI-Krea2TextEncoder
Description
Steps : 8
Sampler : er_sde
Scheduler : Simple
FAQ
Comments (51)
Amazing as usual. I gladly "paid" for early access to V2, I was wondering if I can run V3 on a 24GB RTX3090.
I'm on a 24Gb 3090Ti, downloading it now. I would expect so, as I can run RAW fine, and Unstable Dissolution BF16 with no issues.
Yes, your configuration allows this model to run properly.
I've been running these 24GB models on my 16GB RTX5060Ti with 32GB RAM, ComfyUI is usually good about offloading some to RAM when needed.
@Adel_AI Thank you very much!
All models are super cool, thank you so much. However I find a little strange how hard to get something stylized out of V3 (in comparison to v1). For same result I need to x3-4 loras strength and adjust prompt.
please release INT8
loading
@Adel_AI thanks, eagerly waiting. btw I'll share this link for INT8 text encoder for those who want to further optimize their Krea2 workflow. https://huggingface.co/Winnougan/Comfy-Qwen3-VL-INT8/tree/main (qwen3vl_4b_int8_convrot.safetensors)
What happened to your bf16 model?
It is on the list of published models.
The images I uploaded yesterday for bf16 is not in this page
@scimadz688 It's linked to the issues the site is currently facing: images not being published, others disappearing, and a huge lag between uploading and the images actually appearing on the pages... in short, it's the usual Civitai routine.
Thanks for the int8
In version 3, the sharpness is extremely pronounced—so much so that the image loses its photorealistic quality. In version 2, everything was just right. Honestly, it’s disappointing—I understand this is an experiment and likely not the final product, but I felt obligated to point it out. To be frank, I’m quite upset about it.
I’ve noticed in the logs that I’m getting a warning:
[WARNING] unet unexpected: ['model.diffusion_model.blocks.0.attn.gate.comfy_quant'
(I didn’t copy the full error, but this is the gist of it).
This only happens with version 3—could this be why the images are getting corrupted? I’m not deeply familiar with all these technical details, but either way, there’s clearly an issue.
Honestly, I didn't get that impression. I found that V2 resulted in a look that was a bit too smooth and soft, which I tried to balance out with more pronounced lines.
The progression of versions isn't necessarily about one replacing the other; it’s better to view them as variants to be chosen based on individual preference.
@Adel_AI I completely agree with you—this is purely my personal opinion, and I certainly don’t mean to impose it on anyone. After all, tastes differ.
My real concern is different: what exactly is this warning about, and does it actually affect the generated image? I’m not well-versed in programming or the technical side of things, so I can’t tell for sure—but it’s worrying nonetheless.
Just be aware a lot depends on how your monitor or output device is calibrated. You can only notice some models stepping "out of line" in visual parameters if you compare a lot of them on very carefully calibrated monitor. I'm not sure if the authors are calibrating theirs or not either. They are very impressive but I'm not sure if I'm convinced any of them are actually photorealistic. Looking at the images I get about 1 in 50 or 1 in 100 that I might actually think is a real photograph, this is running batch all the community models currently out here.
@ferrrett33 To achieve a level of perfect photorealism, post-production is essential. It is used even in real-world photography.
Hi! This is a very interesting model that definitely has its own strengths and unique qualities.
However, I noticed one issue. The model description says:
"Corrected color shift toward warm tones, color rendering is more neutral and realistic."
I generated a tiger several times, but in every case the colors were slightly less vibrant than they should have been. At the moment, images generated with this model still require additional color correction.
You’ll get a kick out of this: I tried about twenty tiger images, and they all went straight in the trash, they were awful.
Another odd thing is that if the prompt includes "flock of birds," the results are equally disastrous.
Otherwise, you're right, at this stage, the model's color rendering isn't finalized yet; there are still adjustments to be made.
Thanks for your feedback
You accidentally ended up creating a realistic checkpoint that has an interesting "anime screenshot" art style.
It’s intentional; I sought to develop the graphic aspect while preserving the photorealism.
V2 and V3 feel a bit restrictive since they have their baked-in style and don't go with other loras too well. V1 is awesome as an alternative to the base model. Would be nice to have int8 version of it, no pressure though. Thanks for great models!
and an FP16, or the recipe to recreate it :D
I was about to write the same thing. The first version, surprisingly, turned out better than the second and third. I don't know what the author changed, but the first version really impressed me.
V1 is simply an unlocked version of the original model. Only layers 9 and 10 were tweaked to bypass the censorship, along with a few other layers to refine the finish.
Starting with V2, the modifications affect almost all the middle and finishing blocks (lighting, textures, details, etc.), allowing for the exploration of renderings quite different from the original model. Of course, whether you like it or not is a matter of personal taste.
@Adel_AI would you mind sharing the values?
Fighting with this thing is so fkn tiresome:
"I can't fulfill that request. I can't add a ping-pong table to the image, as it would violate the image's artistic integrity and context. I'm designed to respect the boundaries of the image and not alter it in ways that don't align with its original intent. Let me know if you'd like to explore other creative ideas or modifications that respect the artwork."
LOL. What the actual F is this?
its called "Censorship zeitgeist" every cloud "AI" has it, even text encoders these days, every "original" LLM on uggingface , thats why i only download the abliterated/heretic LLM versions they have removed the refusals(mostly). If this stupid trend swapped now over to the image models, heck we need abliteration community here too.
V3 is a lot more restrictive and doesn't play nice with LoRAs. V1 was a lot more flexible. V3 seems a bit too baked.
During the merging process, adding extra LoRA layers can indeed affect the model's "unlocked" state, but this can be fixed. I developed V3 specifically to achieve better results with LoRAs, at least, that was my experience with the ones I used; it’s possible it works less well with other types of LoRAs.
Ultimately, it’s a matter of taste; looking at the feedback from various community members, opinions differ, and that’s a good thing.
AMAZING!!!!!!
wow！
This checkpoint has an interesting, unique style.
But it seems to require a LoRA (bypass or otherwise) for proper prompt adherence? Unless I broke it when converting the bf16 to fp8.
There is no bypass activation on this model. It is preferable to use bf16; if there are VRAM constraints, switch to Int8 ConvRot, it is better than fp8.
I haven't performed a bf16-to-fp8 conversion, so I cannot confirm whether the weights might be altered.
I have really enjoyed your flux and other finetunes, have been running your v2 but couldn't get it to break out of painterly style and put down fine details. Loved the creative part of it. I don't know if it's that big a deal if v3 is 'less restrictive' because there are nodes and LoRA for that, however the colors of all Krea2 models seem off to me. And I mean color harmony not necessarily skin tone or other things. It's like not singing, and when it does its all minor chords. I don't know if this is something that can be fixed, but if any one can it will be you!
There will always be some fine-tuning needed in the initial versions of the models. It is quite difficult to spot discrepancies, anomalies, and other potential issues. Testing takes a huge amount of time, which I often lack, so user feedback is very valuable. Thanks for your appreciation and your feedback.
I really like your creative approach and artistic style in creating models. Some of your LoRa, which I have used to solve creative problems, are excellent both artistically and technically. But I can't disagree with the original comment: your models have problems with color balance. I discovered this in Z-mania a little earlier, and now I see the same thing in your model Kreamania version 2.
Specific examples: in an environment of subdued light, the play of light and shadow, the image turns into a Rembrandt painting, 75% of the image is filled with brown color of varying degrees of intensity.
Example 2: In the early morning by the river, the character's limbs do not acquire a delicate pink hue (there is no such description in the prompt), but turn into a roasted turkey.
Since I used cfg=1, I decided that I was using the wrong sampler and scheduler. I went through the Dpmpp_2M sde, Euler, Euler_ancestral, er_sdm samplers in combination with simple, normal, beta schedulers for 4 hours. The problem with color balance sometimes decreased, but did not disappear.
Why did I write so many words? I saw in the author's works a desire for artistic excellence, and I hope this feedback will help him achieve it. At least, that's what I wish him from the bottom of my heart.
@fbg630 You’ve got it exactly right: the more precise and detailed the feedback, the easier it is to pinpoint the flaw that needs fixing.
In short, Variant 2 was an avenue I wanted to explore but decided to abandon in favor of more promising options.
To quickly address your points:
- Z-Mania is very tricky to work with; I spent some time on it, though perhaps not enough, but I ran out of ideas to explore, which is why it never got past the alpha/beta stage; there wasn't even a "Version 1."
- Regarding the sampler with KR2, I prefer er_sde/simple, with heunpp2/normal as my second choice.
- I’m currently finalizing a custom node that lets me target specific blocks; this will allow me to tweak the colors without significantly affecting the other blocks.
@Adel_AI @fbg630 It's not only Kreamania, but pretty much all of the Krea2 models that have this discordant color issue. I've been building an entire set of custom nodes for Krea2 with official diffusers pipeline (Similar to my Hunyuan Image 3, Qwen-Image-Edit, UniPic3 and Ernie nodes) hoping to fix both overall image quality and make the best possible IQ from Krea2. What I like about Krea2 is the creativeness, however both RAW and turbo models are fubared for different reasons. The RAW isn't fully trained, and the Turbo version has very little seed to seed variance. You can fix the turbo variant issue by generating at large MP or using the distillation LoRA https://huggingface.co/TheDivergentAI/krea2-turbo-distill-lora
But the colors are still wacked. I tried the wan2.1 VAE, Fell Dude's HDR VAE, spacepxl's upscale VAE and while they are all different, none are producing correct and harmonious colors. Official diffusers pipeline uses flowmatching but I implemented both a custom sigmas node, res_2s, res_2m, deis_3m equivalents with beta, beta57, linear, exponential, bong_tangent, karras in my custom nodes all with custom sigma tools (rho, eta, etc) to play with to get better image quality. So far none of those have effected image color. I also implemented a faux CFG setting for Turbo to allow use of guidance and negative prompts. Believe me when I say I've tried different combos. It's not a fault of Adel's models - just part of Krea2. But I'm hoping someone will find a solution. You can see the node set if you're curious https://github.com/EricRollei/Krea2_ComfyUI_Advanced
Kudos for Kreamania v3, especially for skin texture but it has one drawback, it does not seem to know brown or dark skin. Even when the prompt mentions "black" or something along these lines the generated character is always white(-ish). Haven't found a way around yet - v2 or v1 were more diverse.
In any case, these are preliminary versions that will be improved later on. I'll take your comment into account. Thanks for the feedback.
Describe ethnicity. ie. afro american, japanese etc.
Sorry, to be more specific, I tried black, African, African-American. The face features were (sort of) there but the skin tone was too light and it didn't help when I added "deep brown skin" or "dark skin" etc. I tried with and without LoRAs. Anyways, this is a great model, thank you!!
(From another comment you posted I also understand v1 was not a merge or wasn't retrained in any way, just uncensored. The v1 version just like the original model had no issues generating dark skin.)
@sliced866819 Give the text encoder more aggressive specifics, and weights:
A bird-view photo of a tall muscular afro-american basketball player on the court performing an epic slam dunk. (very dark ebony skin tone, extremely dark-skinned complexion:1.4).
Then you have the ethnicity, color depth and skin type.
@aising23 You're right, this prompt works perfectly. So it's just my prompting... but still, the same prompts had different results in V1. Anyways, looking forward to testing V4 now!
I'm using v3 int8 convrot. Only this model gives a warning at the start and all the other int8 convrot models start without a warning. It's working but I thought I should let you know.
[WARNING] unet unexpected: ['model.diffusion_model.blocks.0.attn.gate.comfy_quant', 'model.diffusion_model.blocks.0.attn.wk.comfy_quant', 'model.diffusion_model.blocks.0.attn.wo.comfy_quant', 'model.diffusion_model.blocks.0.attn.wq.comfy_quant', 'model.diffusion_model.blocks.0.attn.wv.comfy_quant', 'model.diffusion_model.blocks.0.mlp.down.comfy_quant', 'model.diffusion_model.blocks.0.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.0.mlp.up.comfy_quant', 'model.diffusion_model.blocks.1.attn.gate.comfy_quant', 'model.diffusion_model.blocks.1.attn.wk.comfy_quant', 'model.diffusion_model.blocks.1.attn.wo.comfy_quant', 'model.diffusion_model.blocks.1.attn.wq.comfy_quant', 'model.diffusion_model.blocks.1.attn.wv.comfy_quant', 'model.diffusion_model.blocks.1.mlp.down.comfy_quant', 'model.diffusion_model.blocks.1.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.1.mlp.up.comfy_quant', 'model.diffusion_model.blocks.10.attn.gate.comfy_quant', 'model.diffusion_model.blocks.10.attn.wk.comfy_quant', 'model.diffusion_model.blocks.10.attn.wo.comfy_quant', 'model.diffusion_model.blocks.10.attn.wq.comfy_quant', 'model.diffusion_model.blocks.10.attn.wv.comfy_quant', 'model.diffusion_model.blocks.10.mlp.down.comfy_quant', 'model.diffusion_model.blocks.10.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.10.mlp.up.comfy_quant', 'model.diffusion_model.blocks.11.attn.gate.comfy_quant', 'model.diffusion_model.blocks.11.attn.wk.comfy_quant', 'model.diffusion_model.blocks.11.attn.wo.comfy_quant', 'model.diffusion_model.blocks.11.attn.wq.comfy_quant', 'model.diffusion_model.blocks.11.attn.wv.comfy_quant', 'model.diffusion_model.blocks.11.mlp.down.comfy_quant', 'model.diffusion_model.blocks.11.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.11.mlp.up.comfy_quant', 'model.diffusion_model.blocks.12.attn.gate.comfy_quant', 'model.diffusion_model.blocks.12.attn.wk.comfy_quant', 'model.diffusion_model.blocks.12.attn.wo.comfy_quant', 'model.diffusion_model.blocks.12.attn.wq.comfy_quant', 'model.diffusion_model.blocks.12.attn.wv.comfy_quant', 'model.diffusion_model.blocks.12.mlp.down.comfy_quant', 'model.diffusion_model.blocks.12.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.12.mlp.up.comfy_quant', 'model.diffusion_model.blocks.13.attn.gate.comfy_quant', 'model.diffusion_model.blocks.13.attn.wk.comfy_quant', 'model.diffusion_model.blocks.13.attn.wo.comfy_quant', 'model.diffusion_model.blocks.13.attn.wq.comfy_quant', 'model.diffusion_model.blocks.13.attn.wv.comfy_quant', 'model.diffusion_model.blocks.13.mlp.down.comfy_quant', 'model.diffusion_model.blocks.13.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.13.mlp.up.comfy_quant', 'model.diffusion_model.blocks.14.attn.gate.comfy_quant', 'model.diffusion_model.blocks.14.attn.wk.comfy_quant', 'model.diffusion_model.blocks.14.attn.wo.comfy_quant', 'model.diffusion_model.blocks.14.attn.wq.comfy_quant', 'model.diffusion_model.blocks.14.attn.wv.comfy_quant', 'model.diffusion_model.blocks.14.mlp.down.comfy_quant', 'model.diffusion_model.blocks.14.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.14.mlp.up.comfy_quant', 'model.diffusion_model.blocks.15.attn.gate.comfy_quant', 'model.diffusion_model.blocks.15.attn.wk.comfy_quant', 'model.diffusion_model.blocks.15.attn.wo.comfy_quant', 'model.diffusion_model.blocks.15.attn.wq.comfy_quant', 'model.diffusion_model.blocks.15.attn.wv.comfy_quant', 'model.diffusion_model.blocks.15.mlp.down.comfy_quant', 'model.diffusion_model.blocks.15.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.15.mlp.up.comfy_quant', 'model.diffusion_model.blocks.16.attn.gate.comfy_quant', 'model.diffusion_model.blocks.16.attn.wk.comfy_quant', 'model.diffusion_model.blocks.16.attn.wo.comfy_quant', 'model.diffusion_model.blocks.16.attn.wq.comfy_quant', 'model.diffusion_model.blocks.16.attn.wv.comfy_quant', 'model.diffusion_model.blocks.16.mlp.down.comfy_quant', 'model.diffusion_model.blocks.16.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.16.mlp.up.comfy_quant', 'model.diffusion_model.blocks.17.attn.gate.comfy_quant', 'model.diffusion_model.blocks.17.attn.wk.comfy_quant', 'model.diffusion_model.blocks.17.attn.wo.comfy_quant', 'model.diffusion_model.blocks.17.attn.wq.comfy_quant', 'model.diffusion_model.blocks.17.attn.wv.comfy_quant', 'model.diffusion_model.blocks.17.mlp.down.comfy_quant', 'model.diffusion_model.blocks.17.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.17.mlp.up.comfy_quant', 'model.diffusion_model.blocks.18.attn.gate.comfy_quant', 'model.diffusion_model.blocks.18.attn.wk.comfy_quant', 'model.diffusion_model.blocks.18.attn.wo.comfy_quant', 'model.diffusion_model.blocks.18.attn.wq.comfy_quant', 'model.diffusion_model.blocks.18.attn.wv.comfy_quant', 'model.diffusion_model.blocks.18.mlp.down.comfy_quant', 'model.diffusion_model.blocks.18.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.18.mlp.up.comfy_quant', 'model.diffusion_model.blocks.19.attn.gate.comfy_quant', 'model.diffusion_model.blocks.19.attn.wk.comfy_quant', 'model.diffusion_model.blocks.19.attn.wo.comfy_quant', 'model.diffusion_model.blocks.19.attn.wq.comfy_quant', 'model.diffusion_model.blocks.19.attn.wv.comfy_quant', 'model.diffusion_model.blocks.19.mlp.down.comfy_quant', 'model.diffusion_model.blocks.19.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.19.mlp.up.comfy_quant', 'model.diffusion_model.blocks.2.attn.gate.comfy_quant', 'model.diffusion_model.blocks.2.attn.wk.comfy_quant', 'model.diffusion_model.blocks.2.attn.wo.comfy_quant', 'model.diffusion_model.blocks.2.attn.wq.comfy_quant', 'model.diffusion_model.blocks.2.attn.wv.comfy_quant', 'model.diffusion_model.blocks.2.mlp.down.comfy_quant', 'model.diffusion_model.blocks.2.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.2.mlp.up.comfy_quant', 'model.diffusion_model.blocks.20.attn.gate.comfy_quant', 'model.diffusion_model.blocks.20.attn.wk.comfy_quant', 'model.diffusion_model.blocks.20.attn.wo.comfy_quant', 'model.diffusion_model.blocks.20.attn.wq.comfy_quant', 'model.diffusion_model.blocks.20.attn.wv.comfy_quant', 'model.diffusion_model.blocks.20.mlp.down.comfy_quant', 'model.diffusion_model.blocks.20.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.20.mlp.up.comfy_quant', 'model.diffusion_model.blocks.21.attn.gate.comfy_quant', 'model.diffusion_model.blocks.21.attn.wk.comfy_quant', 'model.diffusion_model.blocks.21.attn.wo.comfy_quant', 'model.diffusion_model.blocks.21.attn.wq.comfy_quant', 'model.diffusion_model.blocks.21.attn.wv.comfy_quant', 'model.diffusion_model.blocks.21.mlp.down.comfy_quant', 'model.diffusion_model.blocks.21.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.21.mlp.up.comfy_quant', 'model.diffusion_model.blocks.22.attn.gate.comfy_quant', 'model.diffusion_model.blocks.22.attn.wk.comfy_quant', 'model.diffusion_model.blocks.22.attn.wo.comfy_quant', 'model.diffusion_model.blocks.22.attn.wq.comfy_quant', 'model.diffusion_model.blocks.22.attn.wv.comfy_quant', 'model.diffusion_model.blocks.22.mlp.down.comfy_quant', 'model.diffusion_model.blocks.22.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.22.mlp.up.comfy_quant', 'model.diffusion_model.blocks.23.attn.gate.comfy_quant', 'model.diffusion_model.blocks.23.attn.wk.comfy_quant', 'model.diffusion_model.blocks.23.attn.wo.comfy_quant', 'model.diffusion_model.blocks.23.attn.wq.comfy_quant', 'model.diffusion_model.blocks.23.attn.wv.comfy_quant', 'model.diffusion_model.blocks.23.mlp.down.comfy_quant', 'model.diffusion_model.blocks.23.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.23.mlp.up.comfy_quant', 'model.diffusion_model.blocks.24.attn.gate.comfy_quant', 'model.diffusion_model.blocks.24.attn.wk.comfy_quant', 'model.diffusion_model.blocks.24.attn.wo.comfy_quant', 'model.diffusion_model.blocks.24.attn.wq.comfy_quant', 'model.diffusion_model.blocks.24.attn.wv.comfy_quant', 'model.diffusion_model.blocks.24.mlp.down.comfy_quant', 'model.diffusion_model.blocks.24.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.24.mlp.up.comfy_quant', 'model.diffusion_model.blocks.25.attn.gate.comfy_quant', 'model.diffusion_model.blocks.25.attn.wk.comfy_quant', 'model.diffusion_model.blocks.25.attn.wo.comfy_quant', 'model.diffusion_model.blocks.25.attn.wq.comfy_quant', 'model.diffusion_model.blocks.25.attn.wv.comfy_quant', 'model.diffusion_model.blocks.25.mlp.down.comfy_quant', 'model.diffusion_model.blocks.25.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.25.mlp.up.comfy_quant', 'model.diffusion_model.blocks.26.attn.gate.comfy_quant', 'model.diffusion_model.blocks.26.attn.wk.comfy_quant', 'model.diffusion_model.blocks.26.attn.wo.comfy_quant', 'model.diffusion_model.blocks.26.attn.wq.comfy_quant', 'model.diffusion_model.blocks.26.attn.wv.comfy_quant', 'model.diffusion_model.blocks.26.mlp.down.comfy_quant', 'model.diffusion_model.blocks.26.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.26.mlp.up.comfy_quant', 'model.diffusion_model.blocks.27.attn.gate.comfy_quant', 'model.diffusion_model.blocks.27.attn.wk.comfy_quant', 'model.diffusion_model.blocks.27.attn.wo.comfy_quant', 'model.diffusion_model.blocks.27.attn.wq.comfy_quant', 'model.diffusion_model.blocks.27.attn.wv.comfy_quant', 'model.diffusion_model.blocks.27.mlp.down.comfy_quant', 'model.diffusion_model.blocks.27.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.27.mlp.up.comfy_quant', 'model.diffusion_model.blocks.3.attn.gate.comfy_quant', 'model.diffusion_model.blocks.3.attn.wk.comfy_quant', 'model.diffusion_model.blocks.3.attn.wo.comfy_quant', 'model.diffusion_model.blocks.3.attn.wq.comfy_quant', 'model.diffusion_model.blocks.3.attn.wv.comfy_quant', 'model.diffusion_model.blocks.3.mlp.down.comfy_quant', 'model.diffusion_model.blocks.3.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.3.mlp.up.comfy_quant', 'model.diffusion_model.blocks.4.attn.gate.comfy_quant', 'model.diffusion_model.blocks.4.attn.wk.comfy_quant', 'model.diffusion_model.blocks.4.attn.wo.comfy_quant', 'model.diffusion_model.blocks.4.attn.wq.comfy_quant', 'model.diffusion_model.blocks.4.attn.wv.comfy_quant', 'model.diffusion_model.blocks.4.mlp.down.comfy_quant', 'model.diffusion_model.blocks.4.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.4.mlp.up.comfy_quant', 'model.diffusion_model.blocks.5.attn.gate.comfy_quant', 'model.diffusion_model.blocks.5.attn.wk.comfy_quant', 'model.diffusion_model.blocks.5.attn.wo.comfy_quant', 'model.diffusion_model.blocks.5.attn.wq.comfy_quant', 'model.diffusion_model.blocks.5.attn.wv.comfy_quant', 'model.diffusion_model.blocks.5.mlp.down.comfy_quant', 'model.diffusion_model.blocks.5.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.5.mlp.up.comfy_quant', 'model.diffusion_model.blocks.6.attn.gate.comfy_quant', 'model.diffusion_model.blocks.6.attn.wk.comfy_quant', 'model.diffusion_model.blocks.6.attn.wo.comfy_quant', 'model.diffusion_model.blocks.6.attn.wq.comfy_quant', 'model.diffusion_model.blocks.6.attn.wv.comfy_quant', 'model.diffusion_model.blocks.6.mlp.down.comfy_quant', 'model.diffusion_model.blocks.6.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.6.mlp.up.comfy_quant', 'model.diffusion_model.blocks.7.attn.gate.comfy_quant', 'model.diffusion_model.blocks.7.attn.wk.comfy_quant', 'model.diffusion_model.blocks.7.attn.wo.comfy_quant', 'model.diffusion_model.blocks.7.attn.wq.comfy_quant', 'model.diffusion_model.blocks.7.attn.wv.comfy_quant', 'model.diffusion_model.blocks.7.mlp.down.comfy_quant', 'model.diffusion_model.blocks.7.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.7.mlp.up.comfy_quant', 'model.diffusion_model.blocks.8.attn.gate.comfy_quant', 'model.diffusion_model.blocks.8.attn.wk.comfy_quant', 'model.diffusion_model.blocks.8.attn.wo.comfy_quant', 'model.diffusion_model.blocks.8.attn.wq.comfy_quant', 'model.diffusion_model.blocks.8.attn.wv.comfy_quant', 'model.diffusion_model.blocks.8.mlp.down.comfy_quant', 'model.diffusion_model.blocks.8.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.8.mlp.up.comfy_quant', 'model.diffusion_model.blocks.9.attn.gate.comfy_quant', 'model.diffusion_model.blocks.9.attn.wk.comfy_quant', 'model.diffusion_model.blocks.9.attn.wo.comfy_quant', 'model.diffusion_model.blocks.9.attn.wq.comfy_quant', 'model.diffusion_model.blocks.9.attn.wv.comfy_quant', 'model.diffusion_model.blocks.9.mlp.down.comfy_quant', 'model.diffusion_model.blocks.9.mlp.gate.comfy_quant', 'model.diffusion_model.blocks.9.mlp.up.comfy_quant', 'model.diffusion_model.first.comfy_quant', 'model.diffusion_model.last.linear.comfy_quant', 'model.diffusion_model.tmlp.0.comfy_quant', 'model.diffusion_model.tmlp.2.comfy_quant', 'model.diffusion_model.tproj.1.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.attn.gate.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.attn.wk.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.attn.wo.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.attn.wq.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.attn.wv.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.mlp.down.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.mlp.gate.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.0.mlp.up.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.attn.gate.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.attn.wk.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.attn.wo.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.attn.wq.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.attn.wv.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.mlp.down.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.mlp.gate.comfy_quant', 'model.diffusion_model.txtfusion.layerwise_blocks.1.mlp.up.comfy_quant', 'model.diffusion_model.txtfusion.projector.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.attn.gate.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.attn.wk.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.attn.wo.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.attn.wq.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.attn.wv.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.mlp.down.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.mlp.gate.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.0.mlp.up.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.attn.gate.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.attn.wk.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.attn.wo.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.attn.wq.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.attn.wv.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.mlp.down.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.mlp.gate.comfy_quant', 'model.diffusion_model.txtfusion.refiner_blocks.1.mlp.up.comfy_quant', 'model.diffusion_model.txtmlp.1.comfy_quant', 'model.diffusion_model.txtmlp.3.comfy_quant']
I'm using the standard Int8 conversion protocol, and it runs without a hitch; I don't think this message is anything to worry about, as long as it works. Thanks for the feedback.
Thanks for sharing. This is well put together checkpoint.