Original Title: Pinkstack/LiliPixel-V1-190M
Uploaded by Pinkstack on 11 19, this single-stream DiT was trained from scratch (with 1.91 billion trainable parameters, 24 layers, and a width of 768). It natively outputs pixel art at three resolutions: 16×16, 32×32, and 64×64. The 16 and 32 pixel outputs are RGBA transparent-background sprites, while the 64 pixel output is an opaque scene. Adding "realistic photo," to the beginning of the prompt enables realistic mode. Text encoding uses a frozen MongoDB/voyage-4-nano (approximately 3.46 billion parameters, automatically downloaded on first use), with prompts limited to a maximum of 96 tokens. Training ran for 13.2 ten-thousand steps, using data from four public pixel art datasets.
Why It Matters Game assets can be generated directly as pixel sprites with transparent backgrounds; the author states that inference requires approximately 1GB of memory and can run on CPU, with a ComfyUI node included.
Rising Score: 30 · Likes: 31 · Downloads: 141
Reasons
- The author states that BF16 inference requires approximately 1GB of memory, generating images almost instantly on GPU or even CPU; it is recommended to generate more than 4 images with the same prompt and then select the best one.
- Prompt writing guidelines are clear: end prompts for 16 / 32 pixel outputs with ", transparent background"; for 64 pixel outputs, use the format "subject, action, on <color> background". Shorter prompts are better.
- Includes a single-file ComfyUI custom node and example workflows, capable of outputting RGBA images with alpha channels and masks.
Usage
- diffusers: Use DiffusionPipeline.from_pretrained("Pinkstack/LiliPixel-V1-190M", custom_pipeline="Pinkstack/LiliPixel-V1-190M", trust_remote_code=True). Set height / width to 16, 32, or 64; use 30 steps for resolutions below 32 pixels and 50 steps for 64 pixels, with CFG set to 3.0.
- ComfyUI: Place comfyui-workflow/ComfyUI_LiliPixel.py into custom_nodes, run pip install -U diffusers "transformers>=4.56" accelerate, then drag in LiliPixel_V1.json.
Risks
- With small parameter counts and datasets, complex scenes, rare objects, and text are not rendered well; 16 pixel outputs cannot accommodate more than two subjects, and 64 pixel outputs tend to deviate from the prompt.
- The scene data for 64 pixel outputs was obtained by downscaling synthetic images generated by FLUX; the license is OpenMDW-1.1, requiring trust_remote_code and an internet connection to download the text encoder. This entry has not been tested.