The odd one out in the HobbyLM family: not a language model, but a 333M in-context flow-matching DiT that
generates 1024×1024 images. It was built to see how good a text-to-image model you can train on a genuinely
small budget — the whole thing came together for roughly $300 of Modal GPU time by working in a heavily
compressed latent space instead of pixels.
It runs in the DC-AE f32c32 (SANA-1.1) latent (32× spatial compression → a 32×32×32 latent at 1024px) and
is conditioned on CLIP-L text features, with classifier-free guidance.
Intended use
Text-to-image generation at 1024×1024. Strongest on single objects and cinematic scenes. A sibling 512px
checkpoint additionally does instruction-based image editing.
The two frozen components are not included (download them from their own repos):
mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers (VAE) and openai/clip-vit-large-patch14 (text encoder).
A full from-scratch CPU implementation of this pipeline (CLIP + DiT + DC-AE, in Rust) lives in
hobby-rs.
Samples
1024×1024, generated by this model (CFG ≈ 5, ~100 steps):
HobbyLM-Image scene samples
Results
This is a hobby-scale generator, so the honest "benchmark" is the training curve and qualitative behaviour
rather than FID / GenEval (which we did not compute):
Property
Value
Flow-matching loss (final)
0.76 (lowest of the model lineage — still decreasing)
Parameters
333M (DiT only)
Resolution
1024×1024 (32×32×32 latent)
VAE reconstruction
~26 dB PSNR @512px; sharper at 1024px (32×32 latent)
Qualitatively, the final checkpoint produces accurate objects and cinematic scenes. It is soft on people,
hands, and multi-person scenes — the real small-model / latent-resolution ceiling. Loss was still dropping
at the end of training, so the 333M DiT is not yet saturated.
Files
model.safetensors — the DiT weights.
config.json — DiT config, lat_std, and the VAE scaling_factor.
There is no GGUF build: image-generation DiTs have no standard GGUF runtime.
Limitations
Hands and multi-person scenes are unreliable.
Fine object crispness is capped by the 32× DC-AE latent; a less-compressed VAE would sharpen it at higher cost.
Instruction-based editing is limited (the CLIP-L text encoder is a weak instruction follower); the real
fix is a stronger conditioner, which is future work.
