Skip to content

CHROMA.md

Latest commit

 

History

History
111 lines (78 loc) · 4.65 KB

CHROMA.md

File metadata and controls

111 lines (78 loc) · 4.65 KB

Chroma 1 Quickstart

image

Chroma 1 is an 8.9B-parameter, trimmed variant of Flux.1 Schnell released by Lodestone Labs. This guide walks through configuring SimpleTuner for LoRA training.

Hardware requirements

Despite the smaller parameter count, memory usage is close to Flux Schnell:

  • Quantising the base transformer can still use ≈40–50 GB of system RAM.
  • Rank-16 LoRA training typically consumes:
    • ~28 GB VRAM without base quantisation
    • ~16 GB VRAM with int8 + bf16
    • ~11 GB VRAM with int4 + bf16
    • ~8 GB VRAM with NF4 + bf16
  • Realistic GPU minimum: RTX 3090 / RTX 4090 / L40S class cards or better.
  • Works well on Apple M-series (MPS) for LoRA training, and on AMD ROCm.
  • 80 GB-class accelerators or multi-GPU setups are recommended for full-rank fine-tuning.

Prerequisites

Chroma shares the same runtime expectations as the Flux guide:

  • Python 3.10 – 3.12
  • A supported accelerator backend (CUDA, ROCm, or MPS)

Check your Python version:

python3 --version

Install SimpleTuner (CUDA example):

pip install simpletuner[cuda]

For backend-specific setup details (CUDA, ROCm, Apple), refer to the installation guide.

Launching the web UI

simpletuner server

The UI will be available at http://localhost:8001.

Configuration via CLI

simpletuner configure walks you through the core settings. The key values for Chroma are:

  • model_type: lora
  • model_family: chroma
  • model_flavour: one of
    • base (default, balanced quality)
    • hd (higher fidelity, more compute hungry)
    • flash (fast but unstable – not recommended for production)
  • pretrained_model_name_or_path: leave empty to use the flavour mapping above
  • model_precision: keep default bf16
  • flux_fast_schedule: leave disabled; Chroma has its own adaptive sampling

Example manual configuration snippet

{
  "model_type": "lora",
  "model_family": "chroma",
  "model_flavour": "base",
  "output_dir": "/workspace/chroma-output",
  "network_rank": 16,
  "learning_rate": 2.0e-4,
  "mixed_precision": "bf16",
  "gradient_checkpointing": true,
  "pretrained_model_name_or_path": null
}

⚠️ If Hugging Face access is slow in your region, export HF_ENDPOINT=https://hf-mirror.com before launching.

Dataset & dataloader

Chroma uses the same dataloader format as Flux. Refer to the general tutorial or the web UI tutorial for dataset preparation and prompt libraries.

Training options specific to Chroma

  • flux_lora_target: controls which transformer modules receive LoRA adapters (all, all+ffs, context, tiny, etc.). The defaults mirror Flux and work well for most cases.
  • flux_guidance_mode: constant works well; Chroma does not expose a guidance range.
  • Attention masking is always enabled – ensure your text embedding cache was generated with padding masks (default behaviour in current SimpleTuner releases).
  • Schedule shift options (flow_schedule_shift / flow_schedule_auto_shift) are not needed for Chroma—the helper already boosts tail timesteps automatically.
  • flux_t5_padding: set to zero if you prefer to zero padded tokens before masking.

Automatic tail timestep sampling

Flux used a log-normal schedule that under-sampled high-noise / low-noise extremes. Chroma’s training helper applies a quadratic (σ ↦ σ² / 1-(1-σ)²) remapping to the sampled sigmas so tail regions are visited more often. This requires no extra configuration—it is built into the chroma model family.

Validation & sampling tips

  • validation_guidance_real maps directly to the pipeline’s guidance_scale. Leave it at 1.0 for single-pass sampling, or raise it to 2.0–3.0 if you want classifier-free guidance during validation renders.
  • Use 20 inference steps for quick previews; 28–32 for higher quality.
  • Negative prompts remain optional; the base model is already de-distilled.
  • The model only supports text-to-image at the moment; img2img support will arrive in a later update.

Troubleshooting

  • OOM at startup: enable offload_during_startup or quantise the base model (base_model_precision: int8-quanto).
  • Training diverges early: ensure gradient checkpointing is on, lower learning_rate to 1e-4, and verify captions are diverse.
  • Validation repeats the same pose: lengthen prompts; flow-matching models collapse when prompt variety is low.

For advanced topics—DeepSpeed, FSDP2, evaluation metrics—see the shared guides linked throughout the README.