SolPix

SolPix turns a text prompt into a 512x512 image. Its roughly 49M-parameter flow transformer predicts image latents; Flan-T5 Base encodes the prompt, and the SANA 1.1 DC-AE decodes the result. Both external models stay frozen.

Components

Setting	Value
Generator	Approximately 49M parameters
Architecture	15-block U-shaped joint text/image transformer
Hidden width	512
Attention	8 heads, head dimension 64
FFN	SwiGLU, width 1,152
Long skip connections	7
Conditioning	Shared adaptive layer normalization
Local mixing	Depthwise 3x3 convolution
Objective	Rectified flow matching with logit-normal time sampling
Image latents	32 channels, 32x spatial compression
512x512 latent grid	16x16
Text encoder	Frozen google/flan-t5-base, 768 features, up to 96 tokens
Decoder	Frozen SANA 1.1 DC-AE F32C32
Saved optimizer step	210,000
Configured training schedule	5,000,000 steps

The parameter count covers the generator only. Downloading the frozen text encoder and autoencoder adds separate dependencies. SolPixTransformer2D predicts latent velocity. AutoencoderDCSol loads the corresponding Diffusers AutoencoderDC for decoding.

Generate an image

Download the repository, install its requirements, and give generate.py a prompt and output path:

python -m pip install -r requirements.txt
python generate.py --prompt "A glass greenhouse in a quiet garden after rain" --output solpix.png

To choose a local checkpoint, seed, or sampling settings:

python generate.py \
  --checkpoint ./step_00210000.pt \
  --prompt "A small red sailboat on a misty lake at sunrise" \
  --seed 1234 --steps 40 --guidance-scale 3.5 \
  --output ./solpix.png

On its first run, the helper downloads Flan-T5 Base and the pinned SANA DC-AE revision. It uses Euler integration with classifier-free guidance, running on CUDA when available. CPU inference is supported but slow.

The flow path is x_t = (1 - t) x_clean + t noise. Sampling runs from t=1 down to t=0; encoder and decoder identifiers and revision pins are in config.json.

Data and checkpoint history

We used MONET v1.2.0 with curation seed 20260924. The split contains 174,603 training examples and a 9,300-example validation holdout. SANA F32C32 image latents and Flan-T5 Base caption states were encoded before training.

MONET draws from CC12M, CommonCatalog-CC-BY, COYO, Diffusion-Aesthetic-4K, and LAION. Flux Klein, Flux Schnell, and Z-Image supply synthetic captions. Curation checks resolution, aesthetics, NSFW content, watermarks, and near duplicates.

The source records include CC BY 4.0, Apache 2.0, Google permissive, and MIT license labels. A label on a record doesn't grant a new license to its contents. Images and dataset shards aren't redistributed in this repository.

The Windows v1.0 continuation used BF16 on one RTX 3080 Ti, with batch size 4 and gradient accumulation 16. We released optimizer step 210,000. The documented 1.1 continuation retains that split and targets step 300,000.

Reading the samples

We haven't run a formal image-quality or prompt-following benchmark on this checkpoint. The gallery shows generated examples, without supplying a held-out quality estimate. Composition errors, artifacts, and weak text or fine-detail rendering remain limitations.

There is no built-in safety classifier. Dataset filtering doesn't remove every bias or unwanted association. Flan-T5 and SANA DC-AE have separate licenses and usage terms.

All 15 samples below use the released checkpoint at 512x512, with 32 Euler steps, guidance scale 3.5, and the pinned SANA DC-AE decoder. Their files, prompts, seeds, and SHA-256 values are recorded in samples/.

Sample 01

Prompt: Three Black men sharing french fries at a neighborhood diner, candid documentary photography.
Seed: 260926

Sample 02

Prompt: A red fox standing in fresh snow beneath pine trees at winter dawn, wildlife photography.
Seed: 260927

Sample 03

Prompt: A glass greenhouse filled with ferns after rain, soft natural light, botanical photograph.
Seed: 260928

Sample 04

Prompt: A handmade cobalt blue teapot on a pale stone table, clean studio product photograph.
Seed: 260929

Sample 05

Prompt: A white sailboat crossing a calm blue bay at golden hour, fine art landscape photograph.
Seed: 260930

Sample 06

Prompt: An orange cat curled on a wooden chair in a sunlit bookshop, cozy editorial photograph.
Seed: 260931

Sample 07

Prompt: A small street cafe reflected in wet pavement at night, warm window light, city photograph.
Seed: 260932

Sample 08

Prompt: A wooden lighthouse on a rocky coast under a cloudy sky, atmospheric landscape photograph.
Seed: 260933

Sample 09

Prompt: A bowl of ripe peaches on a kitchen counter, morning light, natural still life photograph.
Seed: 260934

Sample 10

Prompt: A snow-covered cabin among tall pine trees at blue hour, quiet winter landscape photograph.
Seed: 260935

Sample 11

Prompt: A baker placing fresh bread on a cooling rack in a bright kitchen, documentary photograph.
Seed: 260936

Sample 12

Prompt: A goldfinch perched on a thin branch among spring blossoms, close-up wildlife photograph.
Seed: 260937

Sample 13

Prompt: A red bicycle leaning against a brick wall on a leafy neighborhood street, lifestyle photograph.
Seed: 260938

Sample 14

Prompt: A lemon cake with a slice cut out on a ceramic plate, bright tabletop food photograph.
Seed: 260939

Sample 15

Prompt: A small observatory beneath a clear star-filled sky, distant mountains, night landscape photograph.
Seed: 260940

Files

• step_00210000.pt: EMA and raw weights, optimizer state, configuration, and training arguments.
• solpix/: the transformer, decoder adapter, configuration, data, and training components.
• generate.py: prompt-to-image generation. train.py starts training; sample_latents.py samples latents.
• samples/: the 15 PNGs and their metadata. config.json records the architecture and external-model manifest.

License

The code, checkpoint weights, configuration, model card, and supplied banner use Apache 2.0. Attribution is in NOTICE. Dataset, Flan-T5, and SANA DC-AE licenses apply separately.