silveroxides

/

Chroma-GGUF

like 189

Would it really kill the model uploader to respond to this or put a single line of text on the model page explaining what "detail-calibrated" means?
It could literally mean anything. I don't understand why AI devs spend huge amounts of time and compute to make models but writing the worlds most basic text description is beyond them (yes Im kind of annyoed lol)

@awoo621 Here's what Google says. I don't know how accurate this is since its referencing Reddit threads.

The "detail calibrated" version of the Chroma image model is specifically trained to enhance image details, particularly at higher resolutions, while the normal version focuses on broader features and retrieval capabilities like vector search and multi-modal embedding.
Here's a more detailed breakdown:
Normal Chroma Model:
Focus: General image processing, retrieval, and multi-modal embedding.
Features: Vector search, full-text search, document storage, metadata filtering, and multi-modal retrieval.
Embedding: Supports multi-modal embedding functions, allowing it to handle data from different sources (e.g., images and text).
Training: Typically trained on a standard resolution (e.g., 512x512).
Detail Calibrated Chroma Model:
Focus:
Enhancing image details, especially at higher resolutions (e.g., 1024x1024).
Training:
A specialized training approach that prioritizes detail preservation and sharpness.
Benefits:
Potentially produces more realistic and visually appealing images with finer details.
Use Cases:
May be preferred for applications where fine details are crucial, such as creating high-resolution artwork or realistic images.
In essence, the detail calibrated version is a tuned variant of the base Chroma model, optimized for a specific goal – producing images with more intricate detail.

@TotalNoob1 thanks for helping, but tbh as much of an AI fan as I am I'm pretty reluctant to trust the Google AI summaries as in my experience they're often VERY far from the mark and hallucinate tons of info. For starters, I don't believe the "regular" model is training on 512x512 images and the "detail-calibrated" one on 1024x1024. Chroma struggles to produce coherent images at 512px with both versions of the model, and on top of that I doubt they are simultaneously training two models on different resolution datasets. Maybe there is some truth to what Google is saying there, but I'd still guess its wrong on the res and detail-calibrated is better at 2048px resolutions and not 1024px.

It also doesn't match up with my real world experience. I'm mainly using Chroma to make realistic furry art, and it sort of seems to me that the "regular" version produces better details (especially fur) than the "detail-calibrated" version. Maybe the detail-calibrated produces higher quality images for humans only? Who knows lol.

"Flash" versions also just started being added to this repo and ofc there's no explanation for what that means either. If I had to guess, its a low ~2-4 step version of Chroma that works similarly to how base flux Schnell worked. Again, please silveroxides, please just provide the most basic description of what these are...

I appreciate you trying to help though, thank you.