Cinv/cvstudio-illustrator

CV Studio illustrator (on-device)

Ready-to-run conversions used by the CV Studio desktop app to draw one illustration per sentence of a video, on the customer's own machine (no PyTorch at run time):

Folder	Runtime	Contents
macos/	Core ML (macOS 14+, Apple Silicon)	text_encoder.mlpackage (fp16), unet_6bit.mlpackage (6-bit palettised), taesd_decoder.mlpackage, tokenizer.json
windows/	ONNX Runtime (CPU / DirectML)	text_encoder_int8.onnx, unet_int8.onnx (dynamic int8), taesd_decoder.onnx, tokenizer.json

Inputs: UNet (sample 1×4×64×64, timestep 1, encoder_hidden_states 1×77×768) → noise_pred, with the LCM guidance embedding (w = 8) built in; 4 LCM steps; the TAESD decoder returns RGB in [0, 1] at 512×512. manifest.json lists every file with its size and SHA-256.

Origin and licences

• LCM_Dreamshaper_v7 by Simian Luo et al. (https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7), published under the MIT licence; it is a latent consistency model distilled from DreamShaper v7 (Lykon), a Stable Diffusion 1.5 fine-tune released under the CreativeML OpenRAIL-M licence. To respect the upstream terms, the use restrictions of OpenRAIL-M (Attachment A: no illegal, harmful, discriminatory, defamatory or deceptive use, no medical advice, no exploitation of minors, etc.) are passed on to users of these files.
• TAESD (Tiny AutoEncoder for Stable Diffusion) by Ollin Boer Bohan (https://github.com/madebyollin/taesd), MIT.
• The conversions (Core ML palettisation, ONNX export and int8 quantisation) were made by the CV Studio team; no weights were retrained.