Chroma1-HD

Chroma is an 8.9 billion-parameter model based on the FLUX.1-schnell architecture.

Goals and Features:

• Trained on a dataset of 5 million examples (curated from 20 million), including anime, furry art, artistic images, and photographs.
• Addressed content limitations, including the restoration of missing anatomical details.

Architectural Changes:

• Reduced the number of parameters from 12B to 8.9B by removing a layer that redundantly encoded temporal step information (from 3.3B to 250M parameters).
• Utilized the MMDiT masking technique to improve stability and quality.
• Implemented an adapted distribution of training steps to reduce loss spikes.
• Employed the Minibatch Optimal Transport method to accelerate training through vector matching optimization during generation.

Announce Date: 27.01.2025
Parameters: 8.9B
Context: 77
VRAM requirements: 6.9 GB using 4 bits quantization, 13.8 GB using 8 bits quantization, 27.5 GB using 16 bits quantization
Developer: lodestones
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Chroma1-HD capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.