TokForge
Runs on-device in the TokForge app.
TokForge — SD1.5 LCM (MNN)
A Latent Consistency Model (LCM) few-step image generator in MNN format, packaged for fast
on-device CPU image generation in the TokForge Android app (dev.tokforge).
It is a standard Stable Diffusion 1.5 pipeline whose UNet has been LCM-distilled, so it produces a
coherent image in just 4–8 steps with guidance-free sampling (CFG = 1.0) — no separate
negative-prompt pass — which is what makes it fast enough to run on a phone CPU.
Based on
- Base:
Lykon/dreamshaper-7 — an SD1.5-architecture
checkpoint (CLIP text encoder + UNet + VAE).
- LCM adapter:
latent-consistency/lcm-lora-sdv1-5
— the LCM-LoRA consistency adapter, fused into the UNet.
The fused pipeline is exported to the MNN diffusion 3-file layout (text encoder / UNet / VAE decoder).
Format
MNN runtime. The bundle ships the standard diffusion files:
| File | Role |
|---|
unet.mnn (+ .weight) | LCM-distilled UNet |
text_encoder.mnn (+ .weight) | CLIP text encoder |
vae_decoder.mnn (+ .weight) | VAE decoder |
vocab.json, merges.txt, alphas.txt | tokenizer + scheduler data |
Usage
This bundle is loaded automatically by the TokForge Android app — it is not a standalone diffusers
checkpoint. The app fetches the files, loads them through its MNN diffusion engine, and runs few-step
(4–8 step, CFG = 1.0) text-to-image generation on the device CPU.
License & attribution
Released under CreativeML OpenRAIL-M, matching its SD1.5 base.
This model is a derivative of
Lykon/dreamshaper-7
with the
latent-consistency/lcm-lora-sdv1-5
LCM-LoRA adapter. Please retain this attribution and observe the CreativeML OpenRAIL-M use restrictions.
