>I don't fully get why Chroma doesn't get more appreciation.

IMO, the main Chroma's problem is its purpose. In fact, Chroma is FLUX-based analogue of Pony/Illustrious. The vanilla Chroma turn image to anime anytime it sees strange (probably, danbooru tages). It's not easy to do sci fi concepts on vanilla not recieveing anime or semi-realistic output. Chroma has problems with FLUX-trained character LORAs.

And as I can see, attempts to make a photorealism quickly fall it into variation of vanilla FLUX Schnell.

@mphobbit yeah partially agree, usually the culprit is on how sensitive chrome to prompt, since it's also trained on danbooru tags, when i using lot's of tags combined with long ass AI enhanced natural prompt it's still gives semi realistic output, but this Uncanny model help me reduces that issues.

@N3kor0 the same feelings, I spent much time to find the suitable for me combination of sampler. I wait for Radiance release, the test results on current v.0.4. are promising.

Can also recommend GBRX's GonzaLomo Chroma as auxiliary checkpoint, in fact you'll recieve FLUX S. + Chroma two-in-one checkpoint. It also has Chroma's problems with character lora and looks FLUXish but for me it's a second succesful attempt to make photorealistic finetune alongside with UnCanny.

@mphobbit i'am curious on gonzalomo, since may people recommend to use that model, but i don't think my 8gb gpu can handle it, because as far as i know only uncanny model that has quantized version of all chroma finetunes, also this model already gives far better results than base 1hd on some specific scenario.

Agree, idk why my own Lora character that i trained on chroma1 hd still looks so plastic, like half of the time, i think it's maybe because of the training parameters, i hope chroma radiance will give massive boost on realism, and less complicated prompting.

@N3kor0 I have 8 gigs VRAM too :) If you have enough RAM and powerful CPU it hasn't be a serious problem (just need to play with swapping settings, I'm not a Comfy user, on Forge it can be done elementary).

UPD: here is a guide by OGMustard about Comfy Colab notebook for large and heavy checkpoints.

@mphobbit what about ur gen speed? mine with q4 it's around 5 minutes on 35 steps, on 832x1216 or 1k res, i think my ram is sufficient but the generation speed is the problem.

@N3kor0 I don't use ggufs because they are too slow for me (gen itself is about 6-7 mins for 20 steps 1024x1024 but unggufing is slooow) , with FP8 I have around 15 mins for 20 steps on 1024x1024 or 1024x1152 but all this time is spent for generation itself. Unfortunately very rare creators make NF4.

@mphobbit maan 15 min are slow, are you using dpmpp? i got that exact same speed on comfyui using dpmpp, for me it's either euler+beta combo, or res_multistep+bong tangent.

maan 15 min are slow,

@N3kor0 for 8 gigs VRAM it's a miracle that FP8 is generating not for 1-2 hours (and it's somekind of payment for usage almost full-size checkpoint) :) It's not a problem for me to wait (I simply abort the process if see something strange durng gen).

> dpmpp?
No, for Chroma DPM2/Beta (you can see in gallery). Euler/Beta is too FLUXish for me.

Aah, by the way if you use the TA, you can try Gonzalomo Chroma on TA.

@mphobbit i'll try to use fp8, since your gen speed is 15 minutes with fp8, using dpm2/beta, i also got the same 15 minutes gen on gguf, just like dpmpp, so i'll see if res_multistep could give faster results on fp8 models, thanks mate!

@mphobbit  DPM2 doubles the time it takes and along with 20 steps is kinda overkill. The reason why DPM2 produces clean images is that 20 steps = pretty much 40 Steps with DPM2. At a certain step count the model is pretty much done with the rough shape ( 15 steps ) and the rest is just more miniscule details. Euler is terrible for realistic images, dmmpp_2m ( ++ ) is a good middle ground, ddim is ok too but less contrast.

I don't know the options for Forge, but it should have a upscaler somewhere. In some cases you can get away with something like 512 x 768, 768 x 1024 or whatever and upscale it. Even though i can use the full checkpoint and render the picture in my usually 832 x 1280 resolution, it still takes almost 4 minutes without upscale, 768 x 1024 with upscale to 832 x 1344 less then 2 minutes, without much quality loss. If you use the fp8 version that should translate to the same outcome as i with the fp16 version, roughly.

@TijuanaSlumlord thanks for technical details. I tried DPM++ 2M, it didn't impressed me (but some FLUX models I use it). DPM2 was the best for me and gave less anime on vanilla Chroma comparing with Euler/DPM++ 2M (I experimented either locally and on TA).

@dawncreates  I don't get these plastic skin issues that I seeing people talking about. I use clownshark ksampler with res2m + beta57. 15 steps with chroma-flash-heun_r128-fp32 lora + lenovo_chroma lora and/or chroma prof photos lora and/or analog dreams lora. It's about 14s for a standard image (1024x1024) on rtx 5090. Using the right combination of sampler/scheduler and loras just gives consistent realistic photos. No need to write any prompts about the skin. Can even set the CFG as high as 2.5 and get some nice results.

Edit - I didn't even realize that you're Uncanny creator when I wrote that comment! Thanks for the model! Obviously you must be extremely familiar with the in/outs of the model so you don't need to hear my msg lol I just hope more people give it a try. I would love to see more loras get made.

@dawncreates 
> My point was that comma separated tags can steer Chroma-based models in that direction
It's a popular thought but personally I disagree with it. This image was made with NLP but Chroma's general dislike sci fi topics made picture more artistic.

According to my experience, Chroma falls the image towards anime/semi in case if you add something anime or cartoon-related overriding any positives and negatives related to photorealism. For example: this picture doesn't contain a cartoon hero, and this one contains. Both are made with tagging.

@dawncreates The word realistic or anything that has realistic in it, is really bad if you want something "realistic". The model has been trained on a lot of low quality booru stuff, which is nice for the sake of variety, but what people on those anime or furry boorus call realistic or hyper-realistic is everything but. If you go through some of the boorus or fanart, there are a lot of tags that are commonly used and steer the model towards illustrations or cartoon things. Also, i assume a structure like " word1, word2, word3," is more likely to produce something unwanted. Sometimes it works to use tags in between but that can lead to styles mixing, which isn't that good for real images. Things that have more of a real world anchor are better imho, like just photo, candid, analog, found footage, movie still ( that is a double edged sword, there a also cartoon movies, disney pixar stuff ), handheld camera etc ... CGI, render and 3D are also not the best things, because that is general something associated with "realistic".

@TijuanaSlumlord With Flux and its derivatives I don't use the word "realistic" but I do use the word "real" — "real photo", "real photograph"

"And I guess others aren't as concerned about photo realistic male anatomy" It's more difficult to render and the female body gets 95% of the attention, especially with Wan 2.2. I haven't had the best luck with Chroma. My best results, so far, are with Wan 2.2 + loras. Wan 2.2 is lacking in consistency for complex prompts (good genitals for one generation and then nothing is even there the next, the penis only appears for part of the video in the next, only testicles that are too round the next, etc.) but when it gets it right it's more realistic than any Chroma/Flux I have tried. I haven't tried this model, though.

@dawncreates thank you

I usually use the res_multistep/beta57 combination because it is fast and consistent.

@mykeehu thanks

@dawncreates Already tried the ComfyUI template, with no help. Experimented with the suggested sampler/scheduler and cfg/steps. Have to test a different VAE, will do that next. One thing crossed to my mind is there was some kind of glitch while downloading and that messed up something in the model. I'll keep you informed if I figure out the reason. Thanks for the reply!

Update: fp8 and flash are working fine. Re-downloading the fp16.

Update: Looks like the cause of this was a damaged file. After re-downloading the fp16, it worked fine.

@dawncreates Thx brother :)

Most models do not discern age well but will give an approximation that falls into broad categories. Prompting a specific age seldom works because it isn't frequently captioned in training. Uncanny isn't any different. If you want a subject that looks (roughly) 20, use 'late teen adult' in your positive and reinforce desired features like 'smooth skin' and add undesired features to your usual negative, like 'wrinkles, crows feet, gray hair, teen, mature', etc.

And, yes, 'teen' in both the positive and negative does have an affect. The positive pulls the character more toward the youthful side and the negative pushes it away from too youthful. Without it in the negative, you'll get a college freshman; with it you'll get a college senior.

@nyct The exact same prompt works as expected with 1.0. So -

@vekovi8038713dsdsds And training the model on a limited dataset (that does not likely have significant representation of numeric age) dilutes the ability of the model it was trained on... That's pretty much expected. Put in the work to make your prompts work with the model or don't, that's up to you.

Seems to have a middle age women fixation like Flux. Not very useful if you want to generate anime girls IRL.

@dawncreates thank you for the quick and helpful response!

ah, unfortunately I don't have enough vram to load the safetensors. would love it if you could make one to try out, but thanks either way. the images look great!

@dawncreates Would love a GGUF for 1.3 flash, I only have 6 GB VRAM so I am using Q4_K_M for the original chroma flash which fits nicely into VRAM.

@dawncreates thanks!

@dawncreates My bad... I just realized there are two v1.3 models.

Try it with several types of samplers, because prompt tracking didn't work well with all of them for me. For example, it was fast and had prompt tracking with dpmpp_sde or dpmpp_2m_sde.

I couldn't get great upscale results from Chroma 2k or DC, this one works the best for me

BF16 is the full unquantized weights. Make GGUF/FP8/int8 out of it unless you got a lot of vram.

I'd like to know that too. Strangely enough, I have exactly the same problem. The images are pale and noisy, regardless of which sampler/scheduler I use—I've tested many workflows. But I would really like the images themselves. I don't have any problems with GonzaLomo's model, but I don't like the results—everything is too perfect for me.

Are you also using the Flash version? Perhaps there was an error when merging with Flash-Heun-Lora?!

got any advice on how to make it output realistic? My prompting always includes it at the beginning but half the time i get anime.