One ZeroGPU pack per model; each button loads only what it needs

Selective GPU loading + Gemma chat (OAuth) + Florence-2 image / SmolVLM2 video / CLAP music describe tabs; Qwen2.5-VL removed; Qwen3-4B for narration & prompt rewrite; README credits updated

Selective GPU loading + Gemma chat (OAuth) + Florence-2 image / SmolVLM2 video / CLAP music describe tabs; Qwen2.5-VL removed; Qwen3-4B for narration & prompt rewrite; README credits updated

Selective GPU loading + Gemma chat (OAuth) + Florence-2 image / SmolVLM2 video / CLAP music describe tabs; Qwen2.5-VL removed; Qwen3-4B for narration & prompt rewrite; README credits updated

Selective GPU loading + Gemma chat (OAuth) + Florence-2 image / SmolVLM2 video / CLAP music describe tabs; Qwen2.5-VL removed; Qwen3-4B for narration & prompt rewrite; README credits updated

Selective GPU loading + Gemma chat (OAuth) + Florence-2 image / SmolVLM2 video / CLAP music describe tabs; Qwen2.5-VL removed; Qwen3-4B for narration & prompt rewrite; README credits updated

Enhance: cut the chatty opening and Markdown marks from model answers

Run Kokoro in its own process; image request 12 s

Use the bucket: running totals and a Showcase tab

Add task tags and model links so the Space shows up in searches

Image 8 s, enhance image 17 s

Halve every GPU-time estimate

Add torchvision, needed by the vision model

Enhance video: optional music prompt and narration

Add an Enhance tab: describe, rewrite and regenerate images, videos and music

Add a Chat tab (Qwen2.5 7B), which also replaces the small narration planner

Add a Narration tab (Kokoro)

Remove an option this Gradio version does not accept

Add a Transcribe tab (Whisper large v3 turbo)

Add a Music tab (MusicGen medium)

Two tabs (Image, one-button Video), sharper 1152x768 clips, better narration planner

Model cache back on the local disk: the bucket reads too slowly

Keep models on attached storage when it is available

Short description: Kokoro+Qwen narration

Remove the loop cross-fade

Music and narration work alone or together, in any order

Narration button stays locked until music has been added

Name the narration model in the short description

Only the phrase "slow motion" allows slow motion

Avoid slow motion unless the motion text asks for it

Sharper, real-time motion: frame-rate conditioning, motion-first prompt, extended negative prompt

Loop cross-fade lengthened to 2 seconds

Video no longer autoplays

MusicGen medium; narration needs no hint box

Add narration: Qwen2.5 plans the lines and timing, Kokoro speaks them

Add MusicGen large: soundtrack step for the animated clip

LTX-Video 0.9.8 13B distilled on the full GPU, with seamless looping

Cross-fade the end of each clip into its start for a seamless loop

10 fps clips with the first 8 frames trimmed

LTX-Video 0.9.5 with scene+motion prompt and start-image noise, to fix frozen clips

Video step requests 70% of the previous GPU time

Image step requests 10 s of GPU time

Request less GPU time per run so the daily quota lasts longer

Back to the default GPU size; shared T5 encoder keeps it within memory

Use the full GPU (xlarge) for both steps

Share the T5 text encoder between FLUX and LTX-Video to fit the ZeroGPU slice

Remove stray cache

Graphics Studio: FLUX.1-schnell image + LTX-Video animation

initial commit