@Ashen3 I am using natural language prompts, often quite long, but the skin detail leans towards 3d render with your lora compared to just bypass or realism engine

@psspsspsspssspss i find the raw base reacts differently than the turbo to much longer prompts, at least all of my huge ai-generated prompts. I doubt it's maxing out, and may be related to the system's vram. ("Qwen 3-4B has a context length of 32,768 tokens natively"). I find turbo is able to go through the massive prompts without getting bogged down on conflicting or vague terms.

When using this lora, I also get the effect of "plastic" too, I try to fix it by adding other loras with realism, but it doesn't help, since the lora superimposed on top, in addition to quality, pulls over the concept and understanding of the product.

Author, you have amazing work with models, keep up the good work! Thanks for your job!

In the meantime, I will wait for a solution to the plastic problem using a more radical method than "try to add an artistic description".

@Ashen3 I don't really understand why the step counts change on subsequent runs. Weird node. Like my first runs looks like they are applying what the node is saying to calculate. And then every run after that has incorrect values. I mean it still works, just feels ... broken.

@Ashen3 idk i even looked at the source to see if it was obvious. My gut feeling is that since the calculated sigma value is `clone()`d, it feels like something else touches it again. Like you might want to use deepcopy() instead, but idk.

If I hop between different workflows it triggers it pretty consistently. Sometimes the first sampler and second will have 1:1 split, then the next time, the first sampler will only have 1 step and the second will have the rest.

'Krea2 Prompt Weight' node is the solution. You just add (anime:-1) at the end of the prompt and thats it. (my last image have workflow with it included - download it if you like). From KJNodes: https://comfyai.run/documentation/Krea2PromptWeight

I've been tuning a lora stack of lenovo+vhs+realisticsnapshot+galaxyace and more over time. That's been my best defence. But most of the time it happens though it's cause the prompt is just tags, or includes superfluous/conflicting nonsense. Worst case it's cause snofs doesn't have any tagged prompts, but it's more likely that most anime data is described with tags, increasing their normal weight on the base model.

If you aren't using tags, there are inference strategies you can use (to keep the inference noisy as long as possible to induce natural texture). At least adding "raw amateur candid photograph of.... natural lighting" is much less annoying than creating whole bbox's in the magical correct location.

I do get what you're saying about this model in particular. Idk what the data and stuff is, so of course we just get to be vague. I mean I see what you're saying, but there are definitely other "realism" loras I've tried that provide the same "experience" (of higher weights inducing anime styling).

@flo11ok874 I couldn’t quite figure out the link you sent, as I’m using a local ComfyUI installation and couldn’t find that node there. Could you please send me the author’s GitHub link? This node doesn’t seem to be included in the latest version of the KJNodes pack. Thank you.

@darkinlibiti2 It is at KJNodes you just need to have nighty version. https://github.com/kijai/ComfyUI-KJNodes

@Ashen3  try this. It will always have a plastic look to image "A real photograph shot on professional camera of a Japanese woman of East Asian descent with authentic Japanese facial features - monolid or subtle-crease eyes with epicanthic folds, warm undertones, straight brown hair with bangs, fair skin and striking green eyes, slim body, medium breasts and extra large butt standing with hips cocked sharply to one side, back arched, weight on one leg to thrust her massive rear outward, wearing only a sheer micro-thong that vanishes between her cheeks and an unbuttoned cropped shirt hanging open to fully expose her medium breasts and toned midriff, looking back over her shoulder at the viewer with parted lips and heavy-lidded gaze, full-body frontal-three-quarter view from a low angle, soft directional studio lighting that sculpts every curve and shadow. "

Update: surprisingly it actually works fine as long as I don't use the "Krea2 Prompt Weight"

My findings has been if there is anything that reminds it of anime style it will generate plasticky skin. So have to be very careful for the prompt.

@flo11ok874  i setup nightly version? and can't find "Krea2 Prompt Weight" nodes

@Ashen3 it happened on the later versions of Klein 9b also from what i saw.

From my experience btw with krea2 is prefers single resolution training, like only 1024, buckets seem to be fine to use. And there seems to be a hole at 768 where it doesn't like learning. Krea was trained on 256 512 1024.

Instead of wasting your gpu power on multires and adding 1536 you'd probably get sharper results by training with only 1024.

I was training with 512 at first but noticed everything came out blurred and extremely high contrast. As soon as i switched to only 1024 all of my lokr became perfect.

Realism Engine sucks, it fucks up everything. No matter what parameters you use, it makes always do something mediocre

Actually it was caused by ES_SDE with SGM_Uniform, with Euler it works.

Yes

Yes but I recommend using Krea2 Raw with the Turbo lora, I'm getting crazy good results with that and snofs 1.1

@onionhulu781 Any advice on which RAW version to get? I only have 16GB VRAM

@w30357 try krea2_raw_fp8_scaled 12.2gb.

@onionhulu781 what's your sampler/scheduler, steps? lora 0.7 weight?

@w30357 The turbo version is actually a little bigger, not by much but. with the turbo lora you can still generate in 8 steps, but if you decrease strength with the r256 turbo lora I have found down to 0.97 for 10 steps and 0.9 - 0.95 for 12 steps it allows you to go with more steps without burning the image also for refine passes 0.2 on the lora with something like the "FlowMatch Euler Discrete Scheduler (Custom)" node where you can have the exact sigmas you would have at any step, you can run 16 steps on a refine pass from step 36 to 52 and that does a great job at refining, although Krea2 is kinda crap at upscaling it's like the only downside though, but a pass with SeedVR2 to upscale first then a Krea2 pass to refine seems to work really well. Or you can use an upscale model to sharpen the image which work pretty good.

Personally I have an 3 stage workflow which generates the initial image @1.5MP 10 steps, Turbo lora: 0.97, Perturbed-Attention Guidance (Advanced): scale 1.5,then upscales to 6k with 4xNomas8kDAT then downsample with "Scale Image to Total Pixels" node to 2MP with upscale_method: area then feed that image through Krea, 52 steps starting at step 36 ending at step 52—only 16 real steps but allows you to run almost Undistilled therefore you get a quality bump—then upscale with SeedVR2 to 4MP and run USDU with 2 tiles to keep it under Krea2's 2k limit, 8 steps, 0.3 denoise, turbo lora to 0.2.

Also er_sde+simple and node ComfyUI-Krea2T-Enhancer strength: 0.1, for all Krea2 stages.

Basically the propose is to generate the initial image quickly so I can iterate quickly on the initial image to get the composition and most of the broad details that I want, then relay on the 2 latter stages to perfect the image.

If you wanna do this you will need some math nodes to properly scale the tiles for USDU and the shortest edge for SeedVR2.

@Etheoma Thanks for explaining so thoroughly, I really appreciate it! Will have to look into it because when I tried RAW it took so many steps to get something useful that it took way too long to generate.

@w30357 Just btw should update you that I changed to 4x_foolhardy_Remacri.path as an upscaler for the middle stage, as I was having issues with the second pass removing all almost all the skin detail from the first pass, and I found out 4xNomas8kDAT was adding a kind of artifacting to my images which Krea2 was discarding, but also discarding the skin detail from the first pass. 4x_foolhardy_Remacri seems to have a light enough touch that it doesn't create any artifacts, but also heavy duty enough to be useful to de-blurr the image enough for Krea2 to do the rest.

Will be releasing my workflow at some point when I have it perfected, it's already in working order and good enough to be used, but I want to get it working really well.

@Etheoma Thanks again, much appreciated!!!

Try this lora. Works great. Creampie - v1.0 | Krea 2 LoRA | Civitai

@Luxaria The seed on the res node indeed change the output... sheesh genning with two seed nodes?? no thanks.

This thing is definitely a bit too much of a mental challenge.

This helped reduce the anime look. Thanks

@Ca5anovan No problem, glad it helped you! :D

Same for me. Are you using raw or turbo?

Turbo mostly. Specifically raw + turbo lora. I noticed the same result in a workflow where I do half the inference with raw, then the rest with raw + turbo lora. I didn't modify my workflows at all and just dropped in 1.1 instead of 1.0.

1.1 from my experience also messes with the look of named characters in the model, 1.0 didn't have that issue

you need to mention the hair color, then they stay the same. yes its a strange thing it does.

It also auto adds beards and tattoos to men out of nowhere. Hope this gets fixed with an update.