DreamShaper 7 (QNN NPU)

DreamShaper 7, a Stable Diffusion 1.5 fine-tune, compiled for the Qualcomm Hexagon NPU so TokForge can make 512x512 pictures on Snapdragon phones.

Published by TokForge for on-device inference.

What it is NPU image pipeline
Format QNN HTP context binaries, W8A16 (8-bit weights, 16-bit activations), one set per Hexagon generation
Runs on Snapdragon phones (Hexagon NPU), with the CLIP text encoder on the CPU (MNN)
Phone class 8 GB RAM or more (TokForge catalog); Snapdragon 8 Gen 3 or 8 Elite Gen 5
Download 14.7 GB (86 files)
Base model Lykon/dreamshaper-7
License CreativeML OpenRAIL-M plus Lykon's Civitai terms (no selling); the qnn_runtime/ files are under Qualcomm's AI Stack License

TokForge

Runs on-device in the TokForge app.

Use it in TokForge

  1. Install TokForge from Google Play.
  2. On Android, open Models > Recommended > Image and download SD1.5 NPU (Fast). TokForge lists it on phones with 8 GB RAM or more and a supported Snapdragon NPU.
  3. TokForge reads your phone's Hexagon generation and downloads only the matching set from dreamshaper7/ plus the shared CLIP files from clip_sd15_base/ (about 1.2 GB), not the whole repo.
  4. Optional: SD1.5 NPU (Fast, cfg=1) adds the LCM UNet from dreamshaper7-lcm/, which makes pictures in 4 to 8 steps at cfg 1.
  5. Ask for a picture in a chat. The model runs on the phone; once it is downloaded, making pictures needs no connection.

Collections: TokForge Image Models (Android), TokForge NPU Packs.

Files

File Size What it is
aihub/ 3.33 GB in 12 files Qualcomm AI Hub's compiled SD1.5 sets for V75, V79 and V81, with metadata (kept for reference)
clip_sd15_base/ 587 MB in 4 files Shared CPU CLIP text encoder (MNN) with token and position embeddings, plus an ONNX copy of the encoder
dreamshaper7-lcm/ 1.74 GB in 6 files LCM UNet add-on for V75 and V81 (the cfg=1 tier), with its manifest and sub-README
dreamshaper7/ 4.49 GB in 35 files DreamShaper 7 NPU sets for Hexagon V73, V75, V79 and V81 (the files the app downloads), with the manifest and IO contracts
manifest.json 14.3 KB Per-arch file list with sizes and md5 for the base SD1.5 and AI Hub sets
qnn_runtime/ 46.0 MB in 4 files Qualcomm QAIRT 2.40 Hexagon runtime libraries for V75 and V81 (Qualcomm AI Stack License, see Credits)
v73/ 1.12 GB in 6 files Base SD1.5 NPU set for Hexagon V73 (original build, kept for reference)
v75/ 1.12 GB in 6 files Base SD1.5 NPU set for Hexagon V75 (original build, kept for reference)
v79/ 1.12 GB in 6 files Base SD1.5 NPU set for Hexagon V79 (original build, kept for reference)
v81/ 1.12 GB in 6 files Base SD1.5 NPU set for Hexagon V81 (original build, kept for reference)

The app uses dreamshaper7/, dreamshaper7-lcm/ and clip_sd15_base/. The top-level v73/ to v81/ sets with the root manifest.json are the original base SD1.5 build, and aihub/ holds Qualcomm AI Hub's compiled SD1.5; both are kept for reference.

About the model

The shipping checkpoint is DreamShaper-7 (Lykon/dreamshaper-7), an SD1.5 finetune under the CreativeML-OpenRAIL-M license and Lykon's Civitai terms (no selling the model or merges of it). It follows compositional prompts better than base Stable Diffusion 1.5. Same SD1.5 architecture and same temb NPU IO contract as base SD1.5, only the UNet, the VAE-decoder, and the DreamShaper-specific time-embedding MLP weights differ; the CPU CLIP front-end is bit-identical to base SD1.5 and is reused.

The model is quantized to W8A16 (8-bit weights, 16-bit activations) and compiled to QNN HTP context binaries that run image generation directly on the phone's Hexagon DSP. No root is needed.

Architecture Stable Diffusion 1.5 (UNet, VAE decoder, CLIP ViT-L text encoder)
Resolution 512x512 (64x64 latent)
Quantization W8A16, per-channel weights, min-max activation calibration
Compiler Qualcomm AI Runtime (QAIRT) 2.40

What runs where

File (per dreamshaper7/<arch>/ folder) Role
unet.bin UNet HTP context binary (DreamShaper-7 W8A16), on the NPU
vae_decoder.bin VAE decoder HTP context binary, on the NPU
time_mlp.bin Host time-embedding weights (DreamShaper-specific), on the CPU
text_encoder.bin CLIP text-encoder QNN binary (optional; the app uses the CPU CLIP)
tokenizer.json, config.json Tokenizer and pipeline config

The arch-independent CPU CLIP front-end (clip_sd15_base/clip_v2.mnn, token_emb.bin, pos_emb.bin) is shared by every arch and reused across the base + DreamShaper variants.

See dreamshaper7/manifest.json for the authoritative per-arch file set (with per-file size + md5) that the app uses to download the correct binaries for the device.

Chips

These are per-architecture QNN HTP context binaries, one set per Hexagon generation: V73 (Snapdragon 8 Gen 2), V75 (8 Gen 3), V79 (8 Elite) and V81 (8 Elite Gen 5). They are not a portable format like GGUF: each binary is compiled for a specific Hexagon generation. The app reads the device's Hexagon arch and selects the matching set.

Binaries are forward-compatible: a set built for a lower Hexagon arch also runs on a higher-arch DSP, while native-arch sets are preferred for best performance.

TokForge currently offers this NPU route on V75 and V81 phones. On other phones it uses its GPU or CPU image routes. The LCM add-on in dreamshaper7-lcm/ has sets for V75 and V81.

Render settings

Tier Scheduler Guidance Steps UNet file
SD1.5 NPU (Fast) DPM-Solver multistep (diffusers, linspace) cfg 7.5 (two UNet passes per step) 12 by default (up to 20) unet.bin
SD1.5 NPU (Fast, cfg=1) LCM cfg 1 (one UNet pass per step) 4 to 8 unet_lcm.bin

Both tiers render at 512x512 and share the VAE decoder, time-embedding weights, tokenizer and CPU CLIP.

How it was made

  1. Exported Lykon/dreamshaper-7 from diffusers to ONNX: a UNet that takes a precomputed time embedding, and a VAE decoder that takes the raw scheduler latent.
  2. Converted and quantized with QAIRT 2.40 (qairt-converter, qairt-quantizer; W8A16, per-channel weights, min-max activation calibration), then compiled one HTP context binary set per Hexagon arch.
  3. For the dreamshaper7-lcm/ add-on, fused latent-consistency/lcm-lora-sdv1-5 into the UNet first and calibrated along the LCM cfg=1 sampling path.
  4. Converted the CPU CLIP text encoder to MNN.

The aihub/ sets are Qualcomm AI Hub's own compiled Stable Diffusion v1.5 binaries.

Limitations

  • This bundle is loaded automatically by the TokForge Android app; it is not a standalone diffusers checkpoint.
  • Snapdragon only: the binaries need a Qualcomm Hexagon NPU, and TokForge offers the route only on the chips listed above.
  • W8A16 quantization may change pictures slightly compared with the full-precision original.

License

Released under CreativeML OpenRAIL-M, matching its base models.

This model is a derivative of Lykon/dreamshaper-7, itself a finetune of stable-diffusion-v1-5/stable-diffusion-v1-5. Please retain this attribution and observe the CreativeML OpenRAIL-M use restrictions.

DreamShaper 7 is also under Lykon's Civitai terms: no selling the model or merges of it. The LCM-LoRA used for the dreamshaper7-lcm/ add-on is under the CreativeML OpenRAIL++-M license.

The qnn_runtime/ files are Qualcomm software under Qualcomm's AI Stack License, not under the model license (see Credits). TokForge 1.3.9 and later ship these libraries inside the app; earlier versions download them from this folder.

Credits

  • DreamShaper 7 by Lykon (Civitai, Hugging Face). Used under the CreativeML OpenRAIL-M license and the creator's Civitai terms. Lykon's Civitai terms: no selling the model or merges of it.
  • Stable Diffusion v1.5 by Runway and CompVis (Hugging Face). Used under the CreativeML OpenRAIL-M license.
  • LCM-LoRA SD1.5 by Latent Consistency (Hugging Face). Used under the CreativeML OpenRAIL++-M license.
  • Stable-Diffusion-v1.5 (Qualcomm AI Hub) by Qualcomm (Hugging Face). Used under the CreativeML OpenRAIL-M license. The aihub/ sets are its compiled binaries.
  • Compiled for the Hexagon NPU with Qualcomm AI Runtime (QAIRT) 2.40 by Qualcomm.
  • Converted to MNN with MNN by Alibaba (Apache-2.0).
  • The qnn_runtime/ files are Qualcomm AI Runtime (QAIRT) 2.40 runtime libraries by Qualcomm Technologies, Inc. They are Qualcomm software under Qualcomm's AI Stack License (the LICENSE.pdf shipped in the QAIRT SDK), not under the model license above, and are redistributed here for use inside the TokForge app only.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for darkmaniac7/TokForge-SD15-QNN-NPU

Quantized
(2)
this model

Collections including darkmaniac7/TokForge-SD15-QNN-NPU