RealVisXL V4.0 Lightning (GGUF)

RealVisXL V4.0 Lightning, a photoreal SDXL fine-tune, as one GGUF for stable-diffusion.cpp, for making 1024x1024 pictures on high-memory Adreno phones with TokForge.

Published by TokForge for on-device inference.


What it is	Image generator
Format	Single GGUF for stable-diffusion.cpp (q8_0 UNet, f16 VAE, both CLIP text encoders), plus a TAESDXL decoder
Runs on	Android phones with a Qualcomm Adreno GPU (stable-diffusion.cpp, OpenCL)
Phone class	16 GB RAM or more (TokForge catalog)
Download	4.18 GB (4 files)
Base model	SG161222/RealVisXL_V4.0_Lightning
License	CreativeML OpenRAIL++-M plus SG_161222's Civitai terms (credit required, no selling)

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.

Use it in TokForge

Install TokForge from Google Play.
On Android, open Models > Recommended > Image and download RealVisXL Lightning (Adreno Quality). TokForge lists it on phones with 16 GB RAM or more and a Qualcomm Adreno GPU. The download includes the GGUF and the TAESDXL decoder.
Describe the picture you want. The model runs on the phone; once it is downloaded, making pictures needs no connection.

Collections: TokForge GGUF Models, TokForge Image Models (Android).

Files

File	Size	What it is
`manifest.json`	550 B	Render defaults and the GGUF's size and md5
`realvisxl-v40-lightning-q8_0.gguf`	4.18 GB	CLIP-L, OpenCLIP-bigG, Lightning UNet (q8_0) and SDXL VAE (f16) in one GGUF
`taesdxl_decoder_f16.safetensors`	2.45 MB	TAESDXL tiny VAE decoder (f16) for the fast decode option
`MD5SUMS`	68 B	md5 of the GGUF

About the model

A single self-contained GGUF for stable-diffusion.cpp, packaging RealVisXL V4.0 Lightning (SDXL, photoreal, few-step Lightning). The UNet is q8_0, the VAE is kept at f16, both CLIP text encoders included, everything in one file.

This is the GPU "Quality" image tier for the TokForge Android app: native 1024px photoreal SDXL on supported high-memory Qualcomm/Adreno devices via the OpenCL F16-GEMM route. It is not a CPU default (SDXL is too heavy for phone CPU); it is the opt-in photoreal step-up above the DreamShaper fast route.

The small taesdxl_decoder_f16.safetensors is the SDXL tiny VAE decoder (TAESDXL) that TokForge uses for its fast decode option; without it, pictures are decoded with the full SDXL VAE inside the GGUF.

Render settings

SDXL Lightning, few-step:

sampler:      dpm++2m
scheduler:    discrete
steps:        6   (4 = fast floor, 8 = extra refinement)
cfg-scale:    1.5
resolution:   1024x1024 (SDXL native)
backend:      OpenCL on high-memory Adreno (GPU); SDXL CPU is not a supported tier

stable-diffusion.cpp CLI example

sd -M img_gen \
  -m realvisxl-v40-lightning-q8_0.gguf \
  -p "close-up portrait, natural skin texture, 85mm, RAW photo, photorealistic" \
  --sampling-method dpm++2m --steps 6 --cfg-scale 1.5 \
  -W 1024 -H 1024 -o out.png

How it was made

Loaded SG161222/RealVisXL_V4.0_Lightning (SDXL diffusers, fp16). Lightning is a baked few-step distilled checkpoint (no LoRA to fuse).
Exported to a single original-format SDXL .safetensors.
Converted to GGUF with stable-diffusion.cpp (-M convert --type q8_0, keeping the VAE at f16 via --tensor-type-rules '^first_stage_model\.=f16').

Limitations

Built for stable-diffusion.cpp; this is a runtime file, not a diffusers training checkpoint.
TokForge offers it only on 16 GB phones with a Qualcomm Adreno GPU; other phones use the DreamShaper fast route.
q8_0 quantization of the UNet may change pictures slightly compared with the full-precision original.

License

License: openrail++ (inherited from RealVisXL V4.0 Lightning / SDXL).
Base model: RealVisXL V4.0 Lightning by SG161222: https://huggingface.co/SG161222/RealVisXL_V4.0_Lightning
Built on top of Stable Diffusion XL (Stability AI).

The original openrail++ terms and attribution requirements propagate to this GGUF and any images generated with it.

RealVisXL is also under SG_161222's Civitai terms: credit required, no selling the model or merges of it. The TAESDXL decoder is under the MIT license.

Credits

RealVisXL V4.0 Lightning 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.
TAESDXL tiny VAE by madebyollin (Hugging Face). Used under the MIT license.
Converted to GGUF with stable-diffusion.cpp by leejet (MIT).