darkmaniac7/TokForge-SD15-LCM-Photoreal-b1

SD1.5 LCM Photoreal, Realistic Vision V5.1 (MNN, batch-1)

Realistic Vision V5.1 with the LCM-LoRA fused in, as an MNN bundle that TokForge uses to make photoreal images on Android phones.

Published by TokForge for on-device inference.

What it is	Image generator (style: photoreal)
Format	MNN diffusion bundle: batch-1 UNet with 8-bit weights, f16 text encoder and VAE decoder
Runs on	Android phones (CPU)
Phone class	8 GB RAM or more (TokForge catalog)
Download	1.21 GB (9 files)
Base model	SG161222/Realistic_Vision_V5.1_noVAE, latent-consistency/lcm-lora-sdv1-5
License	CreativeML OpenRAIL-M, plus SG_161222's Civitai terms for Realistic Vision (credit, no selling, same permissions)

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

1. Install TokForge from Google Play.
2. On Android, open Models > Recommended > Image and download Photoreal (SD1.5 LCM Realistic Vision). TokForge lists it on phones with 8 GB RAM or more.
3. Choose it as your image model and type a prompt. The image is made on the phone; once the model is downloaded, no connection is needed.

Collections: TokForge Image Models (Android), TokForge SD1.5 LCM Styles (MNN).

Files

File	Size	What it is
alphas.txt	13.3 KB	Scheduler alphas
merges.txt	525 KB	CLIP tokenizer merges
text_encoder.mnn	76.3 MB	CLIP text encoder graph (SD1.5 reference)
text_encoder.mnn.weight	170 MB	CLIP text encoder weights (f16, SD1.5 reference)
unet.mnn	1.04 MB	UNet graph (batch-1 export, dynamic height and width)
unet.mnn.weight	862 MB	UNet weights (8-bit, Realistic Vision V5.1 with LCM fused)
vae_decoder.mnn	154 KB	VAE decoder graph (SD1.5 reference)
vae_decoder.mnn.weight	99.0 MB	VAE decoder weights (f16, SD1.5 reference)
vocab.json	862 KB	CLIP tokenizer vocabulary

Only unet.mnn differs from the plain bundle TokForge-SD15-LCM-Photoreal. Every other file, unet.mnn.weight included, is byte-identical (same sha256).

About the model

Realistic Vision V5.1 by SG_161222 is a photoreal SD1.5 checkpoint. This bundle fuses the LCM-LoRA into its UNet instead of using DreamShaper 7 as the base, so it makes a coherent, photorealistic 512x512 image in 4 steps. It is the Photoreal look in the TokForge image styles: natural skin, lifelike texture and lighting, true-to-life portraits and scenes. Its look is more photographic than the DreamShaper 7 bundles, which lean stylized.

Realistic Vision V5.1 ships "noVAE", so the standard SD1.5 VAE is used (the same VAE as in the DreamShaper bundles).

Architecture	Stable Diffusion 1.5 (CLIP ViT-L/14 text encoder, UNet, VAE decoder)
Base checkpoint	Realistic Vision V5.1 by SG_161222
Few-step adapter	LCM-LoRA SD1.5, fused into the UNet
UNet	8-bit weights (asymmetric), batch 1, dynamic height and width
Text encoder and VAE decoder	SD1.5 reference MNN files, f16
Output	512x512

Render settings

Setting	Value
Steps	4 (TokForge default); up to 8 for extra refinement
CFG scale	1.0 (guidance-free)
Negative prompt	Not used at CFG 1.0
Size	512x512

Batch-1 and plain bundles

This model is published twice. The weights are the same; only the UNet graph differs.

Repo	UNet export	Downloaded by
TokForge-SD15-LCM-Photoreal-b1 (this repo)	batch 1, dynamic height and width (sample=[1,4,-1,-1])	TokForge 3.6.0 beta 24 and later, including the current 1.3.x releases
TokForge-SD15-LCM-Photoreal	fixed batch 2 (conditional and unconditional)	Older TokForge builds, for example 3.6.0 beta 12

LCM sampling is guidance-free (CFG 1.0), so the unconditional UNet pass adds nothing. With the batch-1 UNet, TokForge's engine runs one conditional UNet pass per step instead of a batch of two. Older builds such as TokForge 3.6.0 beta 12 always feed the UNet a batch of two and would crash on the batch-1 export, so they keep downloading the plain repo by its exact name. The two repos are separate so that each app version gets the files it can run.

How it was made

1. In diffusers, load SG161222/Realistic_Vision_V5.1_noVAE, then load and fuse latent-consistency/lcm-lora-sdv1-5 (fuse_lora in fp32).
2. Export the fused UNet to ONNX, traced at batch 1 with dynamic height and width, so MNN reports sample=[1,4,-1,-1].
3. Convert the UNet to MNN with 8-bit weights (command below).
4. Add the SD1.5 reference MNN text encoder and VAE decoder (f16) and the tokenizer and scheduler files. These are not Realistic Vision V5.1's own; they are byte-identical across the TokForge SD1.5 LCM bundles.

MNNConvert -f ONNX --modelFile unet/model.onnx --MNNModel unet.mnn \
  --saveExternalData=1 --weightQuantBits 8 --weightQuantAsymmetric

Modifications (OpenRAIL-M §III "mark modified")

This bundle is a modified derivative of Realistic Vision V5.1:

1. 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.
2. 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.
3. CLIP and VAE are the SD1.5 reference MNN assets, not Realistic Vision's own.

No retraining or fine-tuning of the original weights was performed beyond the LCM fuse + quant.

Limitations

• Built for TokForge's MNN diffusion engine. This is a runtime bundle, not a diffusers checkpoint.
• Needs an engine that can run a batch-1 UNet. Older TokForge builds (for example 3.6.0 beta 12) must use the plain repo.
• Use 4 to 8 steps: below 4 the output is noise, and quality falls off above 8.
• The text encoder and VAE are the SD1.5 reference files, not Realistic Vision V5.1's own.
• The UNet weights are 8-bit, which may slightly change fine detail compared with the full-precision model.

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-license

These use-based restrictions propagate to all who use or redistribute this bundle.

License

This bundle is shared under the CreativeML OpenRAIL-M license; see Use restrictions above. Sources and their licences:

• Base model: SG161222/Realistic_Vision_V5.1_noVAE under CreativeML-OpenRAIL-M (plus the creator's Civitai terms, listed under Credits; redistribution of derivatives is licensed under OpenRAIL-M §III.4). RealisticVision V5.1 by SG_161222.
• LCM adapter: latent-consistency/lcm-lora-sdv1-5 under openrail++ (UNet-only consistency adapter, fused to keep the 4-step floor).

SG_161222's Civitai terms for Realistic Vision V5.1: credit required, no selling the model or merges of it, and derivatives keep the same permissions.

Credits

• Realistic Vision V5.1 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, and derivatives keep the same permissions.
• LCM-LoRA SD1.5 by Latent Consistency (Hugging Face). Used under the CreativeML OpenRAIL++-M license.
• Stable Diffusion v1.5 by Runway and CompVis (Hugging Face). Used under the CreativeML OpenRAIL-M license.
• Converted to MNN with MNN by Alibaba (Apache-2.0).