Chroma1-Base

'An 8.9B text-to-image model built from FLUX.1-schnell, with CLIP removed and the per-block modulation replaced by one small approximator network. Chroma1-Base is the neutral checkpoint meant for fine-tuning; Chroma1-HD shares its architecture and size.'

weights: lodestones/Chroma1-Base ↗
org: lodestones
modality: image
tasks: text-to-image
license: Apache 2.0
released: 2025-07-29not verified
total params: 13.75B
model.arch: chroma

Components

role	model	params	size	dtype	trained
Transformer	Chroma1-Base MMDiT (single file) lodestones/Chroma1-Base/Chroma1-Base.safetensors Chroma (ai-toolkit, original layout) ai-toolkit loads this single file, not the Diffusers transformer/ folder in the same repo (same parameter count). Chroma1-HD.safetensors in lodestones/Chroma1-HD has the same size and layout.	8.90B 8,899,983,424	17.80 GB	bf16	yes
Text encoder	T5-XXL v1.1 (encoder only) lodestones/Chroma1-Base/text_encoder transformers.T5EncoderModel ai-toolkit ignores this folder and loads text_encoder_2/ from ostris/Flex.1-alpha instead (same config, same parameter count and byte size).	4.76B 4,762,310,656	9.52 GB	bf16	no
Tokenizer	T5 SentencePiece (32k vocab) lodestones/Chroma1-Base/tokenizer transformers.T5Tokenizer ai-toolkit loads tokenizer_2/ from ostris/Flex.1-alpha.	—	—	—	—
VAE	FLUX.1 VAE lodestones/Chroma1-Base/vae diffusers.AutoencoderKL ai-toolkit loads vae/ from ostris/Flex.1-alpha, which is the same FLUX.1 VAE (same size and config).	83.8M 83,819,683	168 MB	bf16	no
total		13.75B	27.49 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: FLUX.1 VAE
pixels per token: 16×16
notes: Same latent space as FLUX.1 (scaling 0.3611, shift 0.1159). The transformer has patch_size 1 in its config; the 2×2 patching happens outside it, by packing latents into 64-channel tokens (16 × 2 × 2).

input	latent (c×t×h×w)	tokens
512×512	16×1×64×64	1,024
1024×1024	16×1×128×128	4,096	ai-toolkit sample default
1536×1536	16×1×192×192	9,216

Architecture

Blocks: 19 double-stream + 38 single-stream (FLUX layout)
Hidden size: 3072 (24 heads × 128)
MLP ratio: 4.0
Text conditioning: Joint attention on T5-XXL hidden states (4096-d), 512 tokens, padding masked except one token
Pooled / CLIP input: None. FLUX’s CLIP-L and pooled-vector path are removed
Modulation: An approximator MLP (5 layers, 5120 hidden) maps timestep + a per-slot index to all 344 modulation vectors, replacing FLUX’s per-block modulation layers
Guidance input: Kept in the approximator input but always fed 0 (no distilled guidance)
Objective: Rectified flow
Position: 3-axis RoPE [16, 56, 56], theta 10000

In AI Toolkit

model.arch: chroma
UI label: Chroma (image)
model.name_or_path: lodestones/Chroma1-Base
source: extensions_built_in/diffusion_models/chroma/chroma_model.py
extensions_built_in/diffusion_models/chroma/pipeline.py
extensions_built_in/diffusion_models/chroma/src/model.py
extensions_built_in/diffusion_models/chroma/src/layers.py
toolkit/models/v2/text_encoders/t5.py
toolkit/models/v2/vae/autoencoder_kl.py
toolkit/models/registry.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections

UI defaults

quantize / quantize_te: true / true
noise_scheduler / sampler: flowmatch / flowmatch
network.conv: disabled (linear LoRA only)

Specifics

Checkpoint resolution: name_or_path must resolve to one .safetensors file. lodestones/Chroma picks the newest chroma-unlocked-vNN file, lodestones/Chroma/vNN picks that version, lodestones/Chroma1-<name> downloads <name>.safetensors from that repo; anything else must be a local file.
Text encoder and VAE source: Always ostris/Flex.1-alpha (text_encoder_2/, tokenizer_2/, vae/), not the Chroma repo. The CLIP slot is filled with a dummy module.
Resolution: Buckets snap to multiples of 32.
Approximator: The modulation approximator (distilled_guidance_layer) runs under no_grad in the forward pass, so training only reaches the blocks, not the approximator.
Text encoder training: Not supported. T5 stays frozen.
Loss: Flow-matching target noise − latents. Train scheduler uses dynamic shift (0.5 to 1.15).
Sampling: Flowmatch Euler with resolution-dependent shift. True CFG with the negative prompt runs when guidance_scale > 1.
LoRA target: Chroma
Saving: Full fine-tunes save one file in the original Chroma layout. LoRA keys use the ComfyUI diffusion_model prefix.
Metadata base version: chroma

Example config

not verified

job: extensionconfig:  name: "my_chroma_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 16        linear_alpha: 16      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        train_unet: true        train_text_encoder: false        gradient_checkpointing: true        noise_scheduler: "flowmatch"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "lodestones/Chroma1-Base"        arch: "chroma"        quantize: true        quantize_te: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        neg: ""        guidance_scale: 4        sample_steps: 25        prompts:          - "a bear building a log cabin in the snow covered mountains"