Advanced settings: pick v0.2.1 instead of v0.3 (both LoRAs loaded, switched per call)

v0.3: 6 steps by default, new 9-step setting (7 turbo + 2 base-model steps) (part 2)

v0.3: 6 steps by default, new 9-step setting (7 turbo + 2 base-model steps)

Usage counter: append one metadata-only JSON line per generation to a private dataset (all replicas)

Space title: Viggle Turbo for Qwen-Image-2.1

Prompt enhancement: DeepSeek V4.1 Flash via OpenRouter (off-GPU, Auto/On/Off, Auto under 30 tokens, refs <=0.5 MP, zero-retention providers) (part 2)

Prompt enhancement: DeepSeek V4.1 Flash via OpenRouter (off-GPU, Auto/On/Off, Auto under 30 tokens, refs <=0.5 MP, zero-retention providers)

References: one gallery (drop / + Add image, up to 6, numbered image 1..N) instead of three slots

Result image: Download button back (interactive=False), lossless PNG instead of webp q80

Examples: keep the 8 most striking (adds the promo's character style file); drop the dense-layout claim (part 2)

Examples: keep the 8 most striking (adds the promo's character style file); drop the dense-layout claim

Comparison tab: opens on the transparent anime bride; dense-text examples open on the 8-step turbo

?tab=comparison opens the Comparison tab; model card link above the fold

Comparison tab: drop zh_2 (asks for a third-party character), 32 examples (part 2)

Comparison tab: drop zh_2 (asks for a third-party character), 32 examples

Fix example clicks (cache_examples=False); Custom output size (2048² T2I / 1536² edit cap); UI: short header, Advanced settings, examples reproduce with their seed, Viggle theme (part 2)

Fix example clicks (cache_examples=False); Custom output size (2048² T2I / 1536² edit cap); UI: short header, Advanced settings, examples reproduce with their seed, Viggle theme

Comparison tab (33 Qwen Space examples, turbo 6 steps vs base 40 steps, image slider); 5 examples from the Qwen Space; header: very competitive with the base (part 2)

Comparison tab (33 Qwen Space examples, turbo 6 steps vs base 40 steps, image slider); 5 examples from the Qwen Space; header: very competitive with the base

v0.2.1: default LoRA -> v6_isg step-700 EMA r256, examples re-rendered

v0.2 on 6 steps: raw sigmas 1, .9375, .875, .75, .5, .25 (same weights); steps slider adds steps at the high-noise end only; examples re-rendered

v0.2.1: pencil-sketch example re-rendered with seed 3 (seed 42 left it a colour photo)

v0.2.1: xlarge card, no VAE tiling (tiled reference encode wrecked edits; 3-reference edit OOMed at 44.3 GiB on large)

v0.2.1: VAE tiling on every editing call (3-reference edit OOMed at 44.3 GiB on the 44.7 GiB card)

v0.2: 5-step LoRA r256 student, re-rendered examples

Add 4-step Qwen-Image-2.1 turbo demo

Remove two dropped example results

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Add 4-step Qwen-Image-2.1 turbo demo

Pin torch 2.9.1 + torchvision 0.24.1 (Qwen3VLVideoProcessor needs torchvision)

Add 4-step Qwen-Image-2.1 turbo demo

initial commit