Q-P1NG/qdream-models

Q-Dream models by Q-P1NG

On-device text-to-image models for Android, prepared by Q-P1NG for its Q-Dream engine. 512×512 images in under one second on the Snapdragon 8 Elite NPU (estimated 1.1 s on 8 Gen 3 and 1.3 s on 8 Gen 2 from Qualcomm AI Hub profiles), and offline generation on the CPU of any arm64 phone with 6 GB of RAM or more. No server is involved: prompts and images never leave the phone.

Gallery

Every image below was generated with the CPU bundles of this repository (MNN int8 UNet, TAESD decoder) through the Q-Dream reference pipeline, which uses the same tokenizer, LCM sampler and decoder as the app: 512×512, 6 LCM steps, guidance 1.0, no negative prompt, no upscaling or retouching. The first row uses the same prompt and seed on the three models.

DreamShaper 8 LCM	Realistic Vision V6.0 B1 + LCM-LoRA	CyberRealistic Final + LCM-LoRA

Prompts and seeds

Model	Prompt	Seed
DreamShaper 8 LCM	close-up portrait of an old fisherman with a weathered face and a knitted hat, dramatic light, photo	11
Realistic Vision V6.0 B1	close-up portrait of an old fisherman with a weathered face and a knitted hat, dramatic light, photo	11
CyberRealistic Final	close-up portrait of an old fisherman with a weathered face and a knitted hat, dramatic light, photo	11
DreamShaper 8 LCM	a majestic dragon perched on a cliff above a glowing city at night, epic fantasy, dramatic lighting, highly detailed	11
Realistic Vision V6.0 B1	a lion with a golden mane in the savanna at sunset, wildlife photography, sharp focus	11
CyberRealistic Final	portrait of a young woman with freckles and windswept red hair, golden hour, shallow depth of field, 85mm photo	23
DreamShaper 8 LCM	a magical forest with glowing mushrooms and fireflies at night, fantasy illustration	23
Realistic Vision V6.0 B1	underwater coral reef with tropical fish and sun rays, crystal clear water, vivid colors	11
CyberRealistic Final	a futuristic cyberpunk street at night, neon signs, rain, reflections on wet asphalt, cinematic	11
DreamShaper 8 LCM	a medieval castle on an island at sunrise, fantasy landscape, volumetric light	11
Realistic Vision V6.0 B1	a cozy wooden cabin by a lake under the northern lights, reflections, snow, long exposure photo	23
CyberRealistic Final	an astronaut floating above the earth, sunrise over the planet, stars, photorealistic	11

Bundles

Each folder is a self-contained bundle: models, tokenizer, licenses and a bundle.json manifest with the SHA-256 of every file.

Folder	Model	Runs on	Default	Size
cyberrealistic-final-lcm-npu-v73	CyberRealistic Final + LCM-LoRA (fused)	NPU: Snapdragon 8 Gen 2, 7 Gen 4 (Hexagon v73)	512 px, 6 steps	1.01 GB
cyberrealistic-final-lcm-npu-v75	CyberRealistic Final + LCM-LoRA (fused)	NPU: Snapdragon 8 Gen 3 (Hexagon v75)	512 px, 6 steps	1.01 GB
cyberrealistic-final-lcm-npu-v79	CyberRealistic Final + LCM-LoRA (fused)	NPU: Snapdragon 8 Elite (Hexagon v79)	512 px, 6 steps	1.01 GB
cyberrealistic-final-lcm-npu-v81	CyberRealistic Final + LCM-LoRA (fused)	NPU: Snapdragon 8 Elite Gen 5 (Hexagon v81)	512 px, 6 steps	1.01 GB
dreamshaper8-lcm-npu-v73	DreamShaper 8 LCM	NPU: Snapdragon 8 Gen 2, 7 Gen 4 (Hexagon v73)	512 px, 6 steps	1.01 GB
dreamshaper8-lcm-npu-v75	DreamShaper 8 LCM	NPU: Snapdragon 8 Gen 3 (Hexagon v75)	512 px, 6 steps	1.01 GB
dreamshaper8-lcm-npu-v79	DreamShaper 8 LCM	NPU: Snapdragon 8 Elite (Hexagon v79)	512 px, 6 steps	1.01 GB
dreamshaper8-lcm-npu-v81	DreamShaper 8 LCM	NPU: Snapdragon 8 Elite Gen 5 (Hexagon v81)	512 px, 6 steps	1.01 GB
cyberrealistic-final-lcm-int8	CyberRealistic Final + LCM-LoRA (fused)	CPU, any arm64 phone	384 px, 6 steps	1.10 GB
dreamshaper8-lcm-int8	DreamShaper 8 LCM	CPU, any arm64 phone	384 px, 6 steps	1.10 GB
realisticvision6-lcm-int8	Realistic Vision V6.0 B1 + LCM-LoRA (fused)	CPU, any arm64 phone	384 px, 6 steps	1.10 GB

All bundles use the LCM sampler (4–8 steps, guidance 1–2) and the TAESD decoder; the CPU bundles also include the full SD VAE decoder for a final high-quality decode.

Measured performance

Device	Bundle	Result
Snapdragon 8 Elite, in the Q-P1NG app	NPU v79, 512 px, 6 steps	0.91–0.93 s per image (text encoder 60 ms, UNet 0.70 s, decoder 38 ms); 1.5–1.6 s including engine load and saving; peak memory 1.4 GB
Snapdragon 8 Elite, command-line engine, warm	NPU v79, 512 px, 6 steps	0.76 s
Galaxy S25 / S24 / S23 (Qualcomm AI Hub profile)	NPU v79 / v75 / v73	UNet 124 / 162 / 201 ms per step; TAESD on NPU 8.9 ms (S25)
Galaxy S25 (Qualcomm AI Hub profile)	CyberRealistic Final NPU v79	UNet 122 ms per step, as DreamShaper
Pixel 10a (Tensor G4), command-line engine, warm	CPU int8, 384 px, 6 steps	about 7 s
Pixel 10a, in the Q-P1NG app	CPU int8, 384 px, 6 steps	12–13 s (the app reloads the engine for each image)

The v81 bundles are compiled from the same quantized models but have not been profiled yet.

Fidelity. Each NPU UNet (w8a16: int8 weights, int16 activations) was checked against its FP32 UNet on 8 real LCM sampling inputs on the S23, S24 and S25: cosine similarity 0.998–0.99998. Quantization with 8-bit activations (required by older Hexagon v68/v69 chips) was tested and rejected because it destroys the image, so Snapdragon 888, 778G and 8 Gen 1 use the CPU bundles.

Using the bundles

The bundles are made for the Q-P1NG app, which downloads the right one for the phone and checks every file against bundle.json. They can also be used by other software:

• *.mnn files are MNN models (built with MNN 3.6). Weights are int8, so a recent MNN with low-memory support is recommended.
• unet_npu.bin and taesd_npu.bin are Qualcomm QNN HTP context binaries built with QAIRT 2.50 for the Hexagon version in the folder name; they load only on that generation, with a QNN runtime of version 2.50 or later. Tensor names, layouts and quantization parameters are stored in the binary (for example qnn-context-binary-utility).
• export_meta.json holds the scheduler parameters (SD 1.5, epsilon prediction, scaled-linear betas); tokenizer/ is the CLIP BPE tokenizer.
• Prompts work best in English. The bundles contain no safety checker: applications must add their own content filtering and enforce the use restrictions below.

License

The bundles are released under the Q-P1NG Q-Dream Model License: free use, modification and redistribution, including commercial use, provided that you:

1. include LICENSE.md and NOTICE.md and credit "Q-Dream models by Q-P1NG — https://huggingface.co/Q-P1NG/qdream-models";
2. respect the use-based restrictions of the CreativeML Open RAIL-M and Open RAIL++-M licenses (Attachment A, in licenses/), and pass them on to your users as an enforceable condition;
3. keep the upstream copyright and license notices, and state your own changes.

In short, the restrictions forbid using the models or their outputs for unlawful purposes, to harm minors, to spread false information or personal data to harm others, to defame or harass, to discriminate, for medical advice, or for fully automated decisions that affect people's legal rights. Read the full list in licenses/CreativeML-OpenRAIL-M.txt.

The bundles derive from the Stable Diffusion 1.5 checkpoints listed in NOTICE.md (each credited to its author), LCM-LoRA (latent-consistency), TAESD (Ollin Boer Bohan, MIT) and the CLIP tokenizer (OpenAI, MIT): NOTICE.md gives sources, file hashes and the changes made by Q-P1NG. Q-P1NG claims no rights in the images you generate.

Limitations

Stable Diffusion 1.5 models reproduce the biases of their training data, render text and hands poorly, and can produce inaccurate or offensive content. Few-step LCM sampling trades some detail for speed compared with 20–30 step sampling.