mlx-community/VOSR2-fp32

VOSR2 (CSWRY/VOSR, "vision-only" one-step generative super-resolution, CVPR 2026) converted to MLX in fp32 for the Swift/MLX port xocialize/mlx-vosr-swift (MLXEngine imageUpscale, the generative stills tier). Three components, converted by that repo's oracle/convert_weights.py from the upstream checkpoints, per-tensor exact (no re-quantisation on upload):

file	component	notes
transformer_fp32.safetensors	LightningDiT 1.394 B (VOSR2/checkpoints/ema_model.safetensors)	patch conv (O,kH,kW,I); nn.Sequential children under layers.N; RoPE tables kept
vae_fp32.safetensors	Qwen-Image 2-D VAE (Qwen-Image-vae-2d)	convs (O,kH,kW,I); RMSNorm2D gamma squeezed to (C)
dinov2_fp32.safetensors	DINOv2 ViT-L/14, blocks 0–17 (dinov2_vitl14_pretrain.pth)	only the block the DiT consumes and below; pos_embed pre-interpolated for the 448² input
config.json	the upstream args.json + VAE config + the DINOv2 truncation	key contract in the converter's docstring

Lanes: mlx-community/VOSR2-fp16 is the shipping lane (the port derives its int8 tier from it at load: group-64 weight-only quantisation of the DiT Linears); mlx-community/VOSR2-fp32 is the parity / reference lane. bf16 is not published — measured 37.6 dB against the fp32 reference (the DiT residual stream reaches |126|).

Parity

Against the upstream PyTorch implementation on CPU fp32 (same input, same noise), decoded 512² output: fp32 lane 115.5 dB PSNR (tiled 1024²: 121.9 dB); fp16 lane 49.3 dB; fp16 + int8 DiT 46.7 dB. Details and the per-sub-op table: PORTING-SPEC.md in the port repo.

Use with mlx-vosr-swift

import MLXServeCore, MLXVOSR
let engine = MLXServeEngine()
let id = try await engine.register(VOSRUpscalePackage.registration, configuration: VOSRConfiguration(quant: .fp16))
let out = try await engine.run(ImageUpscaleRequest(image: image, scale: 2), package: id)   // ×1…×4

The engine materializes this repo into its model store on first use (WeightSourcing).

Licences and provenance

Apache-2.0 throughout: VOSR2 (CSWRY/VOSR), the Qwen-Image VAE (Wan 2.1 / Qwen-Image), DINOv2 (Meta). The training corpus of VOSR2 is not disclosed by its authors; this re-host takes the declared weight licence as governing. ⚠️ Generative: it rewrites letters in already-legible small text — route text-heavy inputs through a legibility guard.