TokForge-SDXL-PhotoMaker-GGUF

Published by TokForge for on-device inference.

TokForge

• Website: https://tokforge.ai
• Discord: https://discord.gg/Acv3CBtfVm
• Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
• iOS TestFlight: https://testflight.apple.com/join/jnufjzRr

Runs on-device in the TokForge app.

TokForge: SDXL PhotoMaker (Reference Identity) bundle

The single-subject reference-identity image route for the TokForge Android app. Attach a photo of a person, then render that person in any scene ("me as a superhero, face visible"). PhotoMaker stacks the reference into the img trigger word and merges an ID embedding into the SDXL UNet, so the rendered face stays recognizable while the prompt drives the whole scene.

This bundle runs on the on-device stable-diffusion.cpp engine (TokForge's PhotoMaker port) via the helper's native --photo-maker path. Full SDXL at 1024² is heavy; this tier is offered on 16 GB-class phones. For 8 GB phones, use the lighter darkmaniac7/TokForge-SD15-IPAdapter plus-face tier instead.

Files

MD5SUMS carries integrity hashes.

Why the FULL (non-Lightning) RealVisXL V4.0, and why f16 (not q8)

PhotoMaker conditions identity by merging its ID embedding LoRA-style into the SDXL UNet weights. This needs:

• A multi-step base, not a few-step Lightning checkpoint. PhotoMaker's ID merge takes effect over a start_merge_step schedule across the full ~20-30 step denoise, a 6-step Lightning floor leaves too little room for the identity merge to converge. So this tier ships the full RealVisXL V4.0 (the same base validated in the PhotoMaker identity render-tests), not the Lightning variant the IP-Adapter tier ships.
• An f16 (non-quantized) base. The LoRA-style merge requires F32-promotable weights; on a quantized base the sd.cpp helper asserts (src1 must be F32) on both the OpenCL and CPU backends. The base is therefore kept at f16, same precision/quality choice as the sibling IP-Adapter bundle.

How TokForge uses it

In the app (16 GB+ phones): Image model picker → download "PhotoMaker SDXL (Reference Identity)" → attach a face photo as a reference under chat → prompt the scene. The engine is invoked as:

sd -M img_gen \
   -m realvisxl-v40-fp16.safetensors \
   --photo-maker photomaker-v1.safetensors \
   --input-id-images-dir <dir_with_your_face.jpg> \
   -p "a photo of a man img as a superhero, face visible, detailed face, looking at viewer" \
   -n "<strong negative>" \
   --steps 20 --cfg-scale 5.0 \
   --sampling-method euler -H 1024 -W 1024

PhotoMaker keys identity to the img trigger word, so keep img in the prompt right after the class word (a man img, a woman img). The app inserts this automatically. Keep the face visible and unobstructed for a recognizable identity.

Licenses

This is an aggregate of independently-licensed components, and each retains its own license (the bundled TAESDXL decoder is MIT, see Credits):

• RealVisXL V4.0 base (realvisxl-v40-fp16.safetensors): OpenRAIL++ (SG161222/RealVisXL_V4.0, the SDXL openrail++ license). Use must comply with the OpenRAIL++ use-based restrictions.
• PhotoMaker-v1 adapter (photomaker-v1.safetensors): Apache-2.0 (TencentARC/PhotoMaker).

Provenance

• Base: realvisxl-v40-fp16.safetensors, a self-contained sd.cpp SDXL safetensors built from SG161222/RealVisXL_V4.0 (full, non-Lightning; fp16 variant weights) via the standard diffusers → single-file SDXL converter (convert_diffusers_to_original_sdxl.py --half). Same byte-for-byte layout as the TokForge IP-Adapter tier's SDXL base, full-multi-step weights.
• Adapter: photomaker-v1.safetensors, the sd.cpp-compatible PhotoMaker-v1 safetensors (TencentARC/PhotoMaker, Apache-2.0).

Credits

• RealVisXL V4.0 by SG_161222 (Civitai, Hugging Face). Used under the CreativeML OpenRAIL++-M license and the creator's Civitai terms. SG_161222's Civitai terms: credit required, no selling the model or merges of it.
• Stable Diffusion XL 1.0 by Stability AI (Hugging Face). Used under the CreativeML OpenRAIL++-M license.
• PhotoMaker by Tencent ARC (Hugging Face). Used under the Apache-2.0 license.
• TAESDXL tiny VAE by madebyollin (Hugging Face). Used under the MIT license.
• The single-file SDXL checkpoint was made with convert_diffusers_to_original_sdxl.py from diffusers by Hugging Face (Apache-2.0).
• Runs on-device with stable-diffusion.cpp by leejet (MIT).