IDKger/ideogram4

Fix CSS and remove queue

Dynamic duration: 60s floor (keep 30% margin)

Dynamic GPU duration (per-step model); move remote upsample off-GPU

Update app.py

Update app.py

Update app.py

Add cached gr.Examples (lazy) + generate defaults + fp8 link; remove stale tmp/ and __pycache__/

Release: pip diffusers (PR #13860) + public model ideogram-ai/ideogram-4-nf4; drop bundled diffusers_src

Update app.py

Re-enable AOTI (PR#6 artifact) + model-card blurb + title

Update README.md

Update README.md

RC header: 'first open-weights Ideogram model', 9.3B params; + dynamo-disable warning fix

Disable dynamo (TORCHDYNAMO_DISABLE) to silence outlines_core WON'T CONVERT spam

Drop dequantize (AOTI off -> run nf4 directly)

Bundle exactly PR #6 (df4eb9b) as the diffusers source

Remote Ideogram magic-prompt (default) + local Qwen fallback radio; lazy enhancer; AOTI off (recompiling)

Pre-warm vec-isa probe in parent so aoti_blocks_load is instant per worker (was ~22s)

Move bnb Params4bit.shape fix to a runtime shim in app.py; restore pristine bundled diffusers

Fix bnb Params4bit.shape tuple (.numel via math.prod); unpin bitsandbytes for cu130

Pin ML stack to validated versions (bitsandbytes 0.47.0 fixes Params4bit.shape tuple)

Demo: native pipe.upsample_prompt + spaces.aoti_blocks_load (overlapped) + timing

Bundle PR diffusers (yiyi-refactor-fused + native prompt upsampling)

Apply precompiled AOTI block (dynamic L=16k) to both transformers; ~1.28x on 1024 turbo

Dequant nf4->bf16 in parent/CPU context (persists across ZeroGPU forks) + streamed upsampler progress

bf16 dequant only in warmup (no generate fallback) — surface warmup failures

xlarge GPU + bf16 transformers + upsampling default-on + gr.JSON caption + seed-in-field + non-fatal warmup

Update app.py

Preload upsampler + pipeline on GPU via @spaces.GPU startup warmup

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py

Update requirements.txt

Graft hosted lm_head onto pipe encoder for upsampling (no second model)

Add prompt upsampling (Qwen3-VL+Outlines), 3 modes, v2 checkpoint, Citrus theme: v6.txt

Add prompt upsampling (Qwen3-VL+Outlines), 3 modes, v2 checkpoint, Citrus theme: requirements.txt

Add prompt upsampling (Qwen3-VL+Outlines), 3 modes, v2 checkpoint, Citrus theme: app.py

Rebuild guidance_schedule to match step count

Upload /tmp/ideogram4_space/app.py with huggingface_hub

bnb: math.prod(shape) so tuple works

Drop ./diffusers_src from requirements

Use sys.path for bundled diffusers

Embed diffusers PR source; install locally

Initial Ideogram 4 NF4 demo

initial commit