SolPix — Final Release

A compact latent text-to-image flow transformer from SolIntelligence.

SolPix predicts flow in the latent space of the SANA 1.1 DC-AE. This repository contains its PyTorch implementation, a full training checkpoint, and a small prompt-to-image example that pairs the model with the matching frozen text encoder and decoder.

Release v1.0.0: this published snapshot packages the durable EMA checkpoint at optimizer step 210,000. A separate SolPix 1.1 continuation is training from this checkpoint toward step 300,000. That in-progress run is not represented by the files in this repository; the public weights remain the stable 210,000-step snapshot until a new release is ready.

Model summary


Model	SolPix, approximately 49M parameters
Task	Text-conditioned image generation in latent space
Objective	Rectified flow matching, sampled with a logit-normal time distribution
Image latent	32 channels, 32× spatial compression; 16×16 grid for 512×512 images
Text conditioning	Frozen `google/flan-t5-base` states (768 features, up to 96 tokens)
Decoder	Frozen SANA 1.1 DC-AE F32C32
Published weights	EMA and raw model weights in `step_00210000.pt`
Release	v1.0.0 published snapshot; SolPix 1.1 continuation in progress
Saved optimizer step	210,000 (configured schedule: 5,000,000 steps)

SolPix outputs latent velocity predictions, not RGB pixels. Use the matching DC-AE to decode its sampled latents. The text encoder and autoencoder are external dependencies and are not included in this repository.

Quick start

The included helper downloads this repository's checkpoint, encodes a prompt with Flan-T5 Base, samples with Euler integration and classifier-free guidance, then decodes with the pinned SANA DC-AE revision.

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

For a local checkpoint copy or a different prompt/seed:

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

The helper targets a 512×512 image and defaults to CUDA when available. CPU inference is supported but will be slow. No quality benchmark is claimed for this release.

Architecture

The transformer implementation is solpix.SolPixTransformer2D: a 15-block, 512-wide U-shaped joint text/image transformer. It uses 8 attention heads (64 dimensions each), a 1,152-wide SwiGLU MLP, seven long skip connections, shared adaptive layer-normalization conditioning, and depthwise 3×3 local mixing. The decoder adapter is solpix.AutoencoderDCSol, which loads the external SANA diffusers.AutoencoderDC. The text encoder and image autoencoder remain frozen and outside the approximately 49M-parameter generator.

The model accepts precomputed text states and noisy autoencoder latents. Its sampler follows the training convention x_t = (1 - t) x_clean + t noise, integrating from t=1 down to t=0.

Training data and recipe

The training split was curated from MONET v1.2.0 with a deterministic seed of 20260924. The curation summary records 174,603 training examples and a 9,300-example validation holdout. It used pre-encoded SANA F32C32 latents and captions encoded with Flan-T5 Base; the images and dataset shards are not redistributed in this model repository.

The selected source categories were CC12M, CommonCatalog-CC-BY, COYO, Diffusion-Aesthetic-4K, LAION, and synthetic captions from Flux Klein, Flux Schnell, and Z-Image. Curation applied resolution, aesthetic, NSFW, watermark, and near-duplicate filters. The source metadata recorded the following upstream license labels: CC BY 4.0, Apache 2.0, Google permissive, and MIT. Those labels describe dataset records. The data shards are not included here; their original terms remain separate from this repository's Apache-2.0 license.

The Windows v1.0 continuation used BF16, batch size 4 per device, gradient accumulation 16, and one RTX 3080 Ti. The published checkpoint is from step 210,000. The SolPix 1.1 continuation uses the same curated split and resumes from this checkpoint toward step 300,000. No final validation benchmark was completed for v1.0.

Intended use

Research and experimentation with compact latent text-to-image flow models, sampling behavior, and small-model training. Treat generated images as model outputs requiring human review before publication or consequential use.

Limitations

The published checkpoint is from step 210,000; prompt following and visual quality have not been established by a formal evaluation.
Small capacity and limited training progress may yield weak composition, artifacts, or poor rendering of fine details and text.
The model has no built-in safety classifier or content filter.
The image decoder and text encoder are external, separately licensed components. Their use is governed by their own model terms.
Training-data curation does not remove all biases, errors, or unwanted associations in source captions and images.

Sample gallery

These are 15 individual images generated by the released SolPix checkpoint, not external examples. Each was sampled at 512×512 with 32 Euler steps and guidance scale 3.5 using the pinned SANA DC-AE decoder. The image files, prompts, seeds, and SHA-256 values are in samples/.

Sample 01

Prompt: Three Black men sharing french fries at a neighborhood diner, candid documentary photography.
Checkpoint: step 210000 · Seed: 260926 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 02

Prompt: A red fox standing in fresh snow beneath pine trees at winter dawn, wildlife photography.
Checkpoint: step 210000 · Seed: 260927 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 03

Prompt: A glass greenhouse filled with ferns after rain, soft natural light, botanical photograph.
Checkpoint: step 210000 · Seed: 260928 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 04

Prompt: A handmade cobalt blue teapot on a pale stone table, clean studio product photograph.
Checkpoint: step 210000 · Seed: 260929 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 05

Prompt: A white sailboat crossing a calm blue bay at golden hour, fine art landscape photograph.
Checkpoint: step 210000 · Seed: 260930 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 06

Prompt: An orange cat curled on a wooden chair in a sunlit bookshop, cozy editorial photograph.
Checkpoint: step 210000 · Seed: 260931 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 07

Prompt: A small street cafe reflected in wet pavement at night, warm window light, city photograph.
Checkpoint: step 210000 · Seed: 260932 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 08

Prompt: A wooden lighthouse on a rocky coast under a cloudy sky, atmospheric landscape photograph.
Checkpoint: step 210000 · Seed: 260933 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 09

Prompt: A bowl of ripe peaches on a kitchen counter, morning light, natural still life photograph.
Checkpoint: step 210000 · Seed: 260934 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 10

Prompt: A snow-covered cabin among tall pine trees at blue hour, quiet winter landscape photograph.
Checkpoint: step 210000 · Seed: 260935 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 11

Prompt: A baker placing fresh bread on a cooling rack in a bright kitchen, documentary photograph.
Checkpoint: step 210000 · Seed: 260936 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 12

Prompt: A goldfinch perched on a thin branch among spring blossoms, close-up wildlife photograph.
Checkpoint: step 210000 · Seed: 260937 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 13

Prompt: A red bicycle leaning against a brick wall on a leafy neighborhood street, lifestyle photograph.
Checkpoint: step 210000 · Seed: 260938 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 14

Prompt: A lemon cake with a slice cut out on a ceramic plate, bright tabletop food photograph.
Checkpoint: step 210000 · Seed: 260939 · Sampler: 32 Euler steps · Guidance: 3.5

Sample 15

Prompt: A small observatory beneath a clear star-filled sky, distant mountains, night landscape photograph.
Checkpoint: step 210000 · Seed: 260940 · Sampler: 32 Euler steps · Guidance: 3.5

Repository files

step_00210000.pt — full PyTorch training checkpoint, including EMA and raw model weights, optimizer state, model configuration, and training arguments.
config.json — SolPix architecture and external encoder/decoder manifest.
solpix/ — SolPixTransformer2D, AutoencoderDCSol, configuration, data, and training components. AutoencoderDCSol wraps the pinned upstream Diffusers AutoencoderDC; its decoder weights are not duplicated here.
train.py and sample_latents.py — training and latent-sampling entry points.
generate.py — prompt-to-image example using the pinned external models.
banner.png — Sol Labs / SolPix banner supplied for this release.
samples/ — 15 individual checkpoint-generated PNGs, JSON metadata, and a manifest.
LICENSE and NOTICE — Apache-2.0 license and release attribution.

License

The SolPix source, checkpoint weights, configuration, model card, and supplied banner in this repository are released under the Apache License 2.0. See LICENSE and NOTICE. This license does not relicense upstream datasets or the separately downloaded Flan-T5 and SANA DC-AE models; their own terms apply.

Downloads last month: -

solintellegence
/

SolPix