Search Results
6/14/2025, 7:57:53 AM
>>105588701
Idk anon, I feel like if you don't prompt actual photorealism then Chroma might do its own thing.
For instance here's v35 detailed with a prompt that I had originally tested on V28
>Amateur photograph, a beautiful young Japanese idol woman with short pink hair grasping an alcoholic bottle and showing her soles in her room
https://desu-usergeneratedcontent.xyz/g/image/1746/74/1746749318652.png
I find that Chroma is very literal, so you want to help it do photorealism with a trigger word. The model has consistently excelled at photorealism with that simple prompt, "Amateur photograph". If I switch "Amateur" to "professional" result is very different, not that it's slopped but it's different. That's why it's crucial to test different keywords for it. Now, there may be some prompts where it's slopped anyways because Flux was so slopped, and that's where negs help. This issue has existed since Chroma early days, not an issue necessarily with new training method. You might also want to try just rewording and shortening prompt with an LLM, the fewer words that are slopped, the better (or easier to pick out which word is slopped).
My only issue is that I am not sure if the model is actually improving or not. For instance, notice the fucked fingers, and we are nearing the end. I'm not worried about the capabilities getting worse, because they aren't, but how much different are they now than just taking my v28 gen and rerolling?
Idk anon, I feel like if you don't prompt actual photorealism then Chroma might do its own thing.
For instance here's v35 detailed with a prompt that I had originally tested on V28
>Amateur photograph, a beautiful young Japanese idol woman with short pink hair grasping an alcoholic bottle and showing her soles in her room
https://desu-usergeneratedcontent.xyz/g/image/1746/74/1746749318652.png
I find that Chroma is very literal, so you want to help it do photorealism with a trigger word. The model has consistently excelled at photorealism with that simple prompt, "Amateur photograph". If I switch "Amateur" to "professional" result is very different, not that it's slopped but it's different. That's why it's crucial to test different keywords for it. Now, there may be some prompts where it's slopped anyways because Flux was so slopped, and that's where negs help. This issue has existed since Chroma early days, not an issue necessarily with new training method. You might also want to try just rewording and shortening prompt with an LLM, the fewer words that are slopped, the better (or easier to pick out which word is slopped).
My only issue is that I am not sure if the model is actually improving or not. For instance, notice the fucked fingers, and we are nearing the end. I'm not worried about the capabilities getting worse, because they aren't, but how much different are they now than just taking my v28 gen and rerolling?
Page 1