DreamLite¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
dreamlite-base |
carlofkl/DreamLite-base | a9a0f15 |
CC-BY-NC-4.0 | Verified |
dreamlite-mobile |
carlofkl/DreamLite-mobile | a48c656 |
CC-BY-NC-4.0 | Verified |
Catalog: dreamlite ·
Models: dreamlite-base, dreamlite-mobile · Task: t2i
Architecture. DreamLite combines Qwen3-VL text conditioning, a depthwise-separable U-Net with grouped-query cross-attention, a TinyVAE, and flow-matching Euler sampling. The model uses a horizontally concatenated target/condition latent pair: a blank condition produces an image from text, while an input image supplies the condition for instruction-guided editing.
Checkpoints. dreamlite-base defaults to 30 steps with text guidance 3.5 and image guidance
1.5. dreamlite-mobile is a four-step distilled checkpoint and does not use classifier-free
guidance. Both checkpoints publish bfloat16 safetensors only.
Support. The native text-to-image backend implements the published Qwen3-VL prompt template,
DreamLite U-Net, TinyVAE, dynamic flow-matching schedule, base guidance, and mobile distilled loop.
The reference pipelines also expose image editing; that input mode is not part of this t2i
backend.
Validation. Published tensor headers match strictly, and random-weight component and tiny end-to-end comparisons against both diffusers pipelines are exact. Real-checkpoint parity is covered: all 8 real-checkpoint tests passed with exactly 0 difference in every recorded comparison, and the run took 2 minutes on an M2 Max (CPU) (validation). The catalog records the exact repositories, commits, component files, and bfloat16 precision.
License. The weights are licensed under CC-BY-NC-4.0 and are non-commercial. Review the license before downloading or using them.