AbrahamPJ/nightmare-flashvsr

FlashVSR v1.1 (tiny) for Nightmare Mobile

FlashVSR v1.1 (Junhao Zhuang et al., Apache 2.0) converted to one GGUF in the layout the Nightmare Mobile DiT engine reads (a stable-diffusion.cpp fork, dit/058):

prefix	what	type
dit.*	the one-step streaming Wan 2.1 1.3B DiT (diffusion_pytorch_model_streaming_dmd)	linear weights Q8_0 (or Q4_0), the rest f32
lq.*	the causal LQ projector; each 3-D conv split into one 2-D conv a temporal tap (lq.convN.kT)	f16
dec.decoder.*	TCDecoder; its first conv padded from 784 to 800 input channels (zeros)	f16
context	the fixed prompt embedding (posi_prompt.pth, umT5)	f32

file	size	note
flashvsr_v11_q8_0.gguf	2.18 GB	the one that was tested on the phone: 45.6 dB from the PyTorch reference (PC CPU)
flashvsr_v11_q4_0.gguf	1.47 GB	DiT linears Q4_0: 29.4 dB from the reference, visibly softer. Kept as a record, not recommended

Built for an "Upscale 2x" option on the app's Wan video nodes (a finished clip at twice its width and height, on the phone's NPU in 256 px input tiles with dense attention). That option was not shipped: on a Galaxy S25 Ultra a 2 s 832x480 clip took about 5 minutes to upscale, five times as long as making it. The files stay here for anyone who wants to pick it up. It is a conversion only: the weights and their license are the authors'.