TokForge
Runs on-device in the TokForge app.
TokForge SD1.5-LCM — Cinematic (epiCRealism)
An epiCRealism + LCM SD1.5 MNN bundle — the TokForge SD15_LCM_MNN fast few-step image model,
but with the candid/film-look photoreal epiCRealism checkpoint as the base instead of
DreamShaper-7. Renders a coherent, cinematic-photoreal 512px image in 4 steps on the in-process MNN
diffusion engine — full MNN speed, the film look baked into the UNet weights.
This is the Cinematic tier of the TokForge styled image catalog — the twin photoreal style to
the Photoreal (Realistic Vision V5.1) tier. epiCRealism leans toward a candid, cinematic film look:
naturalistic lighting and colour, true-to-life skin and texture, lifelike people and scenes. It won
the candid/film-look people category in TokForge's internal photoreal-base A/B. Verified on-device
(OnePlus D9500 CPU, 4 steps) rendering a clean candid/film-look portrait at MNN speed.
9-file MNN bundle: unet.mnn(.weight) (INT8, epiCRealism + LCM fused), reference SD1.5
f16 text_encoder.mnn(.weight) + vae_decoder.mnn(.weight), vocab.json, merges.txt,
alphas.txt.
Provenance & licenses
- Base model: emilianJR/epiCRealism
— CreativeML-OpenRAIL-M (commercial-OK; redistribution of derivatives is licensed under
OpenRAIL-M §III.4). epiCRealism is a photoreal/film-look SD1.5 checkpoint.
- LCM adapter: latent-consistency/lcm-lora-sdv1-5
— openrail++ (UNet-only consistency adapter, fused to keep the 4-step floor).
- CLIP / VAE / tokenizer: the standard SD1.5 reference assets (CLIP ViT-L/14 text encoder +
SD1.5 VAE) reused verbatim from the TokForge base LCM bundle. This matches the other TokForge
SD1.5-LCM bundles' 9-file layout and dodges the libDIF CLIP-trace issue with freshly-exported CLIP.
Modifications (OpenRAIL-M §III "mark modified")
This bundle is a modified derivative of epiCRealism:
- The LCM consistency adapter (
lcm-lora-sdv1-5) is fused into the UNet (fp32 ΔW via
diffusers.fuse_lora()) so the model is coherent at 4-8 steps with CFG≈1.0.
- The fused UNet is exported to ONNX and converted to an INT8-quantized MNN model
(
MNNConvert, asymmetric 8-bit weight quant) for on-device CPU inference.
- CLIP and VAE are the SD1.5 reference MNN assets, not epiCRealism's own.
No retraining or fine-tuning of the original weights was performed beyond the LCM fuse + quant.
Use restrictions (OpenRAIL-M Attachment A)
Use of this model is subject to the CreativeML Open RAIL-M license, including the
Attachment A use-based restrictions: you agree not to use the model, or any derivative, to
violate any law; to exploit/harm minors; to generate or disseminate verifiably false information
to harm others; to generate or disseminate personal identifiable information to harm someone; to
defame, disparage or harass others; for fully automated decision-making that adversely affects
legal rights or creates binding obligations; for discrimination or harm to individuals or groups
based on protected characteristics; to exploit vulnerabilities of a specific group; to generate
non-consensual or false content about individuals; or to provide medical advice/interpretation of
medical results as a substitute for professional advice. See the full license text:
https://huggingface.co/spaces/CompVis/stable-diffusion-licenseThese use-based restrictions propagate to all who use or redistribute this bundle.
Attribution
- epiCRealism — © emilianJR, CreativeML-OpenRAIL-M.
- Latent Consistency Model LoRA (SD1.5) — Latent Consistency team, openrail++.
- Packaged for on-device MNN inference by TokForge (dev.tokforge).
