Upload mmh3-raven-streaming-q4tp.cmf with huggingface_hub

0.5.94: the q8_2f build runs on the card — 357.3 s to 171.5 s

the q8_2f build's host-path cost, measured

Upload mmh3-turbo-clipproj4b-fl2va-v2-q8_2f.cmf.sha256 with huggingface_hub

the two-field eight-bit build: q8_2f, 26.90 GB, cortiq verify clean

0.5.93: the latent upscaler, the q8_2f build, reference clips

Upload images/upscaler_lanczos_1024.png with huggingface_hub

Upload images/upscaler_net_1024.png with huggingface_hub

the ported latent upscaler: 512x288 render decoded at 1024x576

the vision-row ClipProj twin fitted for the v2 build, so the container can be rebuilt

the adapter's cost re-measured on a machine with headroom: 1.06x, which is the noise floor

adapter A/B still: lora

adapter A/B still: base

adapter A/B sample: routed

adapter A/B sample: lora

adapter A/B sample: base

0.5.92: runtime adapters, the branch router, and what this port does not do yet

card: field notes from discussions #1/#2/#4 — 24 GB frame budgets, keyframe VAE encode on CPU, kill disarmed during prompt encode (0.5.80), P-core QoS

clipproj4b-fl2va v2: split projection — vision rows through a vision-fitted map, faces hold

clipproj4b-fl2va v2: split projection — vision rows through a vision-fitted map, faces hold

images: 4x face-crop A/B panels, clipproj v1 vs v2

images: 4x face-crop A/B panels, clipproj v1 vs v2

images: 4x face-crop A/B panels, clipproj v1 vs v2

images: 4x face-crop A/B panels, clipproj v1 vs v2

images: clipproj v1 vs v2 face A/B frames

images: clipproj v1 vs v2 face A/B frames

card: the compact fl2va file — start/end frames on 16-24 GB machines

clipproj + fl2va in one 14.5 GB file: the 4B tower and the VAE encoder join the compact build; first-frame conditioning verified frame-for-frame

README: lead with the three-way size comparison, decide by VRAM up front

README: one 'Making it smaller' section, a Which-file table keyed on VRAM, three-way sample comparison

ClipProj section: scope the timings to the card they were taken on

ClipProj 4B encoder variant: 23.47 GB -> 13.16 GB, 2x faster, still four bits everywhere

ClipProj 4B prompt encoder: 25.20 GB -> 14.13 GB, everything still q4tp

samples: ClipProj 4B variant, same prompt and seed as the q4tp/q2tp A-B

samples: ClipProj 4B variant, same prompt and seed as the q4tp/q2tp A-B

samples: ClipProj 4B variant, same prompt and seed as the q4tp/q2tp A-B

add the host-only row: 646.1 s against 60.2 on the card, same machine and seed

0.5.68: measured on the current build — 60.2 s at four steps, per-stage table, and how to use two cards

0.5.68: the render is on the card end to end — 91.6 s at 8 steps, and a seed finally reproduces

card: name the library that reads this file

Upload README.md with huggingface_hub

Upload mmh3-turbo-fl2va-q2tp.cmf with huggingface_hub

Upload README.md with huggingface_hub

Upload samples/ab_q4tp.mp4 with huggingface_hub

Upload samples/ab_q2tp.mp4 with huggingface_hub

Upload samples/ab_q2tp.gif with huggingface_hub

Upload samples/ab_q4tp.gif with huggingface_hub

Upload README.md with huggingface_hub

Upload mmh3-turbo-fl2va-q4tp.cmf with huggingface_hub

Upload samples/i2v_keyframe.ppm with huggingface_hub