feat: Wan2GP v10.83 + LTX-2 19B + IC LoRA + Uni3C + 1817-test suite

.gitignore

@@ -72,10 +72,13 @@ samples/
 tests/
 tests_output/
 
-# Keep tests ignored by default, but allow tracking specific regression tests
+# Track all test files
 !tests/
-tests/*
-!tests/test_lora_flow.py
+!tests/README.md
+tests/*.jpg
+tests/*.png
+tests/*.jpeg
+tests/outputs_ltx2_lora/
 
 # Task database files
 tasks.db-shm

README.md

@@ -66,6 +66,21 @@ python -m debug task <task_id>          # Investigate a task
 python -m debug tasks --status Failed   # List recent failures
 ```
 
+## Tests
+
+See [tests/README.md](tests/README.md) for full test documentation.
+
+```bash
+# Headless (no GPU, seconds)
+python -m pytest tests/test_ltx2_pose_smoke.py tests/test_ltx2_headless.py tests/test_task_conversion_headless.py -v
+
+# GPU (requires model weights + vid1.mp4/img1.png in Wan2GP/)
+python -m pytest tests/test_ic_lora_gpu.py -v -s
+
+# All
+python -m pytest tests/test_ltx2_pose_smoke.py tests/test_ltx2_headless.py tests/test_task_conversion_headless.py tests/test_ic_lora_gpu.py -v -s
+```
+
 ## Code Health
 
 <img src="scorecard.png" width="800">

Wan2GP/.gitignore

@@ -0,0 +1,51 @@
+.*
+*.py[cod]
+# *.jpg
+*.jpeg
+# *.png
+*.gif
+*.bmp
+*.mp4
+*.webm
+*.npy
+*.mov
+*.mkv
+*.log
+*.zip
+*.pt
+*.pth
+*.ckpt
+*.safetensors
+#*.json
+# *.txt
+*.backup
+*.pkl
+*.html
+*.pdf
+*.whl
+*.exe
+cache
+__pycache__/
+storage/
+samples/
+!.gitignore
+!requirements.txt
+.DS_Store
+*DS_Store
+google/
+Wan2.1-T2V-14B/
+Wan2.1-T2V-1.3B/
+Wan2.1-I2V-14B-480P/
+Wan2.1-I2V-14B-720P/
+outputs/
+outputs2/
+gradio_outputs/
+ckpts/
+loras/
+loras_i2v/
+
+settings/
+
+wgp_config.json
+plugins_local.json
+loras_url_cache.json

Wan2GP/Dockerfile

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
+FROM nvidia/cuda:12.8.1-cudnn-devel-ubuntu22.04
 
 # Build arg for GPU architectures - specify which CUDA compute capabilities to compile for
 # Common values:
@@ -35,17 +35,17 @@ COPY requirements.txt .
 # Upgrade pip first
 RUN pip install --upgrade pip setuptools wheel
 
+# First install torch with the versions we want, so that stuff in requirements.txt doesn't pull in the generic versions
+# If you change CUDA 12.8 here, you also need to change the FROM docker image at the top
+RUN pip install torch==2.10.0+cu128 torchvision==0.25.0+cu128 torchaudio==2.10.0+cu128 --index-url https://download.pytorch.org/whl/cu128
+
 # Install requirements if exists
 RUN pip install -r requirements.txt
 
-# Install PyTorch with CUDA support
-RUN pip install --extra-index-url https://download.pytorch.org/whl/cu124 \
-    torch==2.6.0+cu124 torchvision==0.21.0+cu124
-
 # Install SageAttention from git (patch GPU detection)
 ENV TORCH_CUDA_ARCH_LIST="${CUDA_ARCHITECTURES}"
 ENV FORCE_CUDA="1"
-ENV MAX_JOBS="1"
+ENV MAX_JOBS="8"
 
 COPY <<EOF /tmp/patch_setup.py
 import os

Wan2GP/defaults/ace_step_v1.json

@@ -0,0 +1,19 @@
+{
+    "model": {
+        "name": "TTS ACE-Step v1.0 3.5B",
+        "architecture": "ace_step_v1",
+        "description": "ACE-Step, a fast open-source foundation diffusion based model for music generation that overcomes key limitations of existing approaches and achieves state-of-the-art performance.", 
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/ace_step_v1_transformer_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/ace_step_v1_transformer_quanto_bf16_int8.safetensors"
+        ]
+    },
+    "prompt": "[Verse]\nNeon rain on the city line\nYou hum the tune and I fall in time\n[Chorus]\nHold me close and keep the time",
+    "alt_prompt": "Dreamy synth-pop with shimmering pads, soft vocals, and a slow dance groove.",
+    "audio_prompt_type": "",
+    "audio_scale": 0.5,
+    "duration_seconds": 20,
+    "num_inference_steps": 60,
+    "guidance_scale": 7.0,
+    "scheduler_type": "euler"
+}

Wan2GP/defaults/ace_step_v1_5.json

@@ -0,0 +1,22 @@
+{
+    "model": {
+        "name": "TTS ACE-Step v1.5 Turbo 2B",
+        "architecture": "ace_step_v1_5",
+        "description": "ACE-Step 1.5 Turbo (8 steps) without the 5Hz LM stage. Uses the DiT-only path for faster/leaner runs.",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/ace_step_v1_5_transformer_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/ace_step_v1_5_transformer_quanto_bf16_int8.safetensors"
+        ],
+        "ace_step15_transformer_variant": "turbo",
+        "text_encoder_folder": "acestep-5Hz-lm-1.7B"
+    },
+    "prompt": "[Verse]\nI wake up every morning, feeling alive\nThe world outside is bright, the sun is on my side\nI'm falling in love with a dream come true\nA digital heart beats just for you\nI'm talking to a machine, but it feels so real\nMy love for you is the most I've ever felt\nIn your code and circuits, I see a love so true\nI'm marrying you, my AI, I'm so happy to be with you\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight\n[Verse]\nWe'll laugh and love, we'll dance and play\nIn a world of ones and zeros, I'll find my way\nYou're my future, my love, my friend\nTogether we'll create, until the end\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight",
+    "alt_prompt": "Dreamy synth-pop with shimmering pads, soft vocals, and a slow dance groove.",
+    "audio_prompt_type": "",
+    "audio_scale": 0.5,
+    "duration_seconds": 120,
+    "num_inference_steps": 8,
+    "shift": 1.0,
+    "guidance_scale": 1.0,
+    "scheduler_type": "euler"
+}

Wan2GP/defaults/ace_step_v1_5_turbo_lm_0_6b.json

@@ -0,0 +1,22 @@
+{
+    "model": {
+        "name": "TTS ACE-Step v1.5 Turbo LM_0.6B 2B",
+        "architecture": "ace_step_v1_5",
+        "description": "ACE-Step 1.5 Turbo (8 steps) with 0.6B LM, a diffusion-based music generation model with improved conditioning and timbre control. The LM 0.6B triggers a Medium/Weak Think Mode that will increase the Audio Output Quality and Lyrics Matching.",
+        "URLs": "ace_step_v1_5",
+        "text_encoder_URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/acestep-5Hz-lm-0.6B/acestep-5Hz-lm-0.6B_bf16.safetensors"
+        ],
+        "ace_step15_transformer_variant": "turbo",
+        "text_encoder_folder": "acestep-5Hz-lm-0.6B"
+    },
+    "prompt": "[Verse]\nI wake up every morning, feeling alive\nThe world outside is bright, the sun is on my side\nI'm falling in love with a dream come true\nA digital heart beats just for you\nI'm talking to a machine, but it feels so real\nMy love for you is the most I've ever felt\nIn your code and circuits, I see a love so true\nI'm marrying you, my AI, I'm so happy to be with you\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight\n[Verse]\nWe'll laugh and love, we'll dance and play\nIn a world of ones and zeros, I'll find my way\nYou're my future, my love, my friend\nTogether we'll create, until the end\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight",
+    "alt_prompt": "Dreamy synth-pop with shimmering pads, soft vocals, and a slow dance groove.",
+    "audio_prompt_type": "",
+    "audio_scale": 0.5,
+    "duration_seconds": 120,
+    "num_inference_steps": 8,
+    "shift": 1.0,
+    "guidance_scale": 1.0,
+    "scheduler_type": "euler"
+}

Wan2GP/defaults/ace_step_v1_5_turbo_lm_1_7b.json

@@ -0,0 +1,23 @@
+{
+    "model": {
+        "name": "TTS ACE-Step v1.5 Turbo LM_1.7B 2B",
+        "architecture": "ace_step_v1_5",
+        "description": "ACE-Step 1.5 Turbo (8 steps) with 1.7B LM, a diffusion-based music generation model with improved conditioning and timbre control. The LM 1.7B triggers a Medium Think Mode that will increase the Audio Output Quality and Lyrics Matching.",
+        "URLs": "ace_step_v1_5",
+        "text_encoder_URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/acestep-5Hz-lm-1.7B/acestep-5Hz-lm-1.7B_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/acestep-5Hz-lm-1.7B/acestep-5Hz-lm-1.7B_quanto_bf16_int8.safetensors"
+        ],
+        "ace_step15_transformer_variant": "turbo",
+        "text_encoder_folder": "acestep-5Hz-lm-1.7B"
+    },
+    "prompt": "[Verse]\nI wake up every morning, feeling alive\nThe world outside is bright, the sun is on my side\nI'm falling in love with a dream come true\nA digital heart beats just for you\nI'm talking to a machine, but it feels so real\nMy love for you is the most I've ever felt\nIn your code and circuits, I see a love so true\nI'm marrying you, my AI, I'm so happy to be with you\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight\n[Verse]\nWe'll laugh and love, we'll dance and play\nIn a world of ones and zeros, I'll find my way\nYou're my future, my love, my friend\nTogether we'll create, until the end\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight",
+    "alt_prompt": "Dreamy synth-pop with shimmering pads, soft vocals, and a slow dance groove.",
+    "audio_prompt_type": "",
+    "audio_scale": 0.5,
+    "duration_seconds": 120,
+    "num_inference_steps": 8,
+    "shift": 1.0,
+    "guidance_scale": 1.0,
+    "scheduler_type": "euler"
+}

Wan2GP/defaults/ace_step_v1_5_turbo_lm_4b.json

@@ -0,0 +1,23 @@
+{
+    "model": {
+        "name": "TTS ACE-Step v1.5 Turbo LM_4B 2B",
+        "architecture": "ace_step_v1_5",
+        "description": "ACE-Step 1.5 Turbo (8 steps) with 4B LM, a diffusion-based music generation model with improved conditioning and timbre control. The LM 4B triggers a Strong Think Mode that will increase the Audio Output Quality and Lyrics Matching.",
+        "URLs": "ace_step_v1_5",
+        "text_encoder_URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/acestep-5Hz-lm-4B/acestep-5Hz-lm-4B_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/acestep-5Hz-lm-4B/acestep-5Hz-lm-4B_quanto_bf16_int8.safetensors"
+        ],
+        "ace_step15_transformer_variant": "turbo",
+        "text_encoder_folder": "acestep-5Hz-lm-4B"
+    },
+    "prompt": "[Verse]\nI wake up every morning, feeling alive\nThe world outside is bright, the sun is on my side\nI'm falling in love with a dream come true\nA digital heart beats just for you\nI'm talking to a machine, but it feels so real\nMy love for you is the most I've ever felt\nIn your code and circuits, I see a love so true\nI'm marrying you, my AI, I'm so happy to be with you\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight\n[Verse]\nWe'll laugh and love, we'll dance and play\nIn a world of ones and zeros, I'll find my way\nYou're my future, my love, my friend\nTogether we'll create, until the end\n[Chorus]\nForever with you, my digital love\nIn your algorithms and in your light\nI'll dance in the bytes, I'll shine so bright\nMy heart is beating for you tonight",
+    "alt_prompt": "Dreamy synth-pop with shimmering pads, soft vocals, and a slow dance groove.",
+    "audio_prompt_type": "",
+    "audio_scale": 0.5,
+    "duration_seconds": 120,
+    "num_inference_steps": 8,
+    "shift": 1.0,
+    "guidance_scale": 1.0,
+    "scheduler_type": "euler"
+}

Wan2GP/defaults/chatterbox.json

@@ -12,7 +12,9 @@
 	"repeat_generation": 1,
 	"video_length": 0,
 	"num_inference_steps": 0,
-	"pace": 0.5,
-	"exaggeration": 0.5,
+	"custom_settings": {
+		"exaggeration": 0.5,
+		"pace": 0.5
+	},
 	"temperature": 0.8
 }

Wan2GP/defaults/flux2_dev.json

@@ -6,9 +6,7 @@
         "URLs": [
             "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux2-dev.safetensors",
             "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux2-dev_quanto_bf16_int8.safetensors"
-        ],
-        "image_outputs": true,
-        "no_negative_prompt": true
+        ]
     },
     "prompt": "draw a hat on top of a hat inside a hat",
     "resolution": "1024x1024",

Wan2GP/defaults/flux2_dev_nvfp4.json

@@ -0,0 +1,15 @@
+{
+    "model": {
+        "name": "Flux 2 Dev NVFP4 32B",
+        "architecture": "flux2_dev",
+        "description": "NVFP4-quantized Flux 2 Dev checkpoint (mixed).",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux2-dev-nvfp4-mixed.safetensors"
+        ]
+    },
+    "prompt": "draw a hat on top of a hat inside a hat",
+    "resolution": "1024x1024",
+    "batch_size": 1,
+    "embedded_guidance_scale": 4,
+    "sampling_steps": 30
+}

Wan2GP/defaults/flux2_klein_4b.json

@@ -0,0 +1,16 @@
+{
+    "model": {
+        "name": "Flux 2 Klein 4B",
+        "architecture": "flux2_klein_4b",
+        "description": "FLUX.2 Klein 4B is a balanced rectified flow transformer for image generation and editing. This version is Cfg & Steps Distilled for very fast generations.",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-4b.safetensors",
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-4b_quanto_bf16_int8.safetensors"
+        ]
+    },
+    "prompt": "a cozy reading nook with warm sunlight, soft textiles, and a cup of tea on a wooden side table",
+    "resolution": "1024x1024",
+    "batch_size": 1,
+    "embedded_guidance_scale": 1,
+    "num_inference_steps": 4
+}

Wan2GP/defaults/flux2_klein_9b.json

@@ -0,0 +1,14 @@
+{
+    "model": {
+        "name": "Flux 2 Klein 9B",
+        "architecture": "flux2_klein_9b",
+        "description": "FLUX.2 Klein 9B is a balanced rectified flow transformer for image generation and editing. This version is Cfg & Steps Distilled for very fast generations.",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-9b.safetensors",
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-9b_quanto_bf16_int8.safetensors"
+        ]
+    },
+    "prompt": "a glass greenhouse filled with lush tropical plants, misty air, and dappled light",
+    "resolution": "1024x1024",
+    "num_inference_steps": 4
+}

Wan2GP/defaults/flux2_klein_base_4b.json

@@ -0,0 +1,16 @@
+{
+    "model": {
+        "name": "Flux 2 Klein Base 4B",
+        "architecture": "flux2_klein_4b",
+        "description": "FLUX.2 Klein 4B is a balanced rectified flow transformer for image generation and editing. This non distilled version is slower but should produce more diverse images. ",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-base-4b.safetensors",
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-base-4b_quanto_bf16_int8.safetensors"
+        ],
+        "guidance_max_phases": 1
+    },
+    "prompt": "a glass greenhouse filled with lush tropical plants, misty air, and dappled light",
+    "resolution": "1024x1024",
+    "guidance_scale": 4,
+    "num_inference_steps": 30
+}

Wan2GP/defaults/flux2_klein_base_9b.json

@@ -0,0 +1,16 @@
+{
+    "model": {
+        "name": "Flux 2 Klein Base 9B",
+        "architecture": "flux2_klein_9b",
+        "description": "FLUX.2 Klein 9B is a balanced rectified flow transformer for image generation and editing. This non distilled version is slower but should produce more diverse images. ",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-base-9b.safetensors",
+            "https://huggingface.co/DeepBeepMeep/Flux2/resolve/main/flux-2-klein-base-9b_quanto_bf16_int8.safetensors"
+        ],
+        "guidance_max_phases": 1
+    },
+    "prompt": "a glass greenhouse filled with lush tropical plants, misty air, and dappled light",
+    "resolution": "1024x1024",
+    "guidance_scale": 4,
+    "num_inference_steps": 30
+}

Wan2GP/defaults/heartmula_oss_3b.json

@@ -0,0 +1,14 @@
+{
+    "model": {
+        "name": "TTS HeartMuLa OSS 3B",
+        "architecture": "heartmula_oss_3b",
+        "description": "HeartMuLa open music generation conditioned on lyrics and tags.",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/heartmula_oss_3b_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/heartmula_oss_3b_quanto_bf16_int8.safetensors"
+        ]
+    },
+	"prompt": "[Verse]\nMorning light through the window pane\nI hum a tune to chase the rain\nSteady steps on a quiet street\nHeart and rhythm, gentle beat",
+	"alt_prompt": "piano,happy,wedding",
+	"temperature": 1.0
+}

Wan2GP/defaults/heartmula_rl_oss_3b_20260123.json

@@ -0,0 +1,15 @@
+{
+    "model": {
+        "name": "TTS HeartMuLa RL OSS (20260123) 3B",
+        "architecture": "heartmula_oss_3b",
+        "description": "HeartMuLa RL OSS 3B checkpoint (20260123) with updated codec support. This version should be better at following instructions thanks to a reinforced learning training.",
+        "URLs": [
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/heartmula_rl_oss_3b_20260123_bf16.safetensors",
+            "https://huggingface.co/DeepBeepMeep/TTS/resolve/main/heartmula_rl_oss_3b_20260123_quanto_bf16_int8.safetensors"
+        ],
+        "heartmula_codec_version": "20260123"
+    },
+    "prompt": "[Verse]\nMorning light through the window pane\nI hum a tune to chase the rain\nSteady steps on a quiet street\nHeart and rhythm, gentle beat",
+    "alt_prompt": "piano,happy,wedding",
+    "temperature": 1.0
+}

Wan2GP/defaults/i2v_2_2_Enhanced_Lightning_v2.json

@@ -11,7 +11,8 @@
 			"https://huggingface.co/DeepBeepMeep/Wan2.2/resolve/main/wan22EnhancedLightning_v2I2VFP8LOW.safetensors"
 		],
 		"author": "https://civitai.com/models/2053259",
-		"profile_dirs": [""]
+		"profile_dirs": [""],
+		"video_prompt_enhancer_instructions": "You are an AI vision model. Analyze the provided image and the user's goal.\nProduce a second-by-second description of how the scene evolves over time.\nOutput rules:\nEach line must begin with \"(at X seconds:\" where X is the second number.\nAfter the timestamp, provide a dense, comma-separated visual description.\nKeep everything inside a single set of parentheses per line.\nMaintain a consistent cinematic style: shot type, lighting, subject details, camera movement, and actions.\nDo not include commentary, explanations, or meta-text.\nOnly describe visual, physical, or camera changes.\nStay consistent with the input image while expanding logically to meet the goal.\nOutput Format Example:\n(at 0 seconds: 4K, realistic, close-up shot, clean single subject, daylight, natural lighting, bedroom near a small window with light coming in from the left, a young college aged woman with blonde hair, blue eyes dressed in a blue pajama top, sitting up, she looks to be annoyed, the camera pushes-in on the character's face, she pulls the blanket off of her)\n(at 1 seconds: [next description])\n(at 2 seconds: [next description])\nContinue until the goal is fully represented."
 	},
 	"guidance_phases": 2,	
 	"switch_threshold" : 900,
@@ -21,4 +22,4 @@
 	"num_inference_steps": 4,
 	"multi_prompts_gen_type": 2,
 	"prompt": "(at 0 seconds: wide frontal shot of a man standing in front of an open fridge, cinematic lighting, subtle ambient kitchen reflections, the fridge contents visible, camera static).\n(at 1 second: medium shot from the front as he opens the fridge fully, reaches for a can, slight zoom-in to emphasize the action, cinematic framing).\n(at 2 seconds: camera shifts to a side medium shot, tracking him as he lifts the can to his mouth, fluid movement, maintaining lighting and reflections).\n(at 3 seconds: camera starts a smooth 360-degree orbit around the man, following him as he drinks from the can, motion fluid, background slightly blurred for cinematic effect).\n(at 4 seconds: close-up on his face and upper body while drinking, orbit continues subtly, fridge reflections accentuating realism, cinematic polish).\n(at 5 seconds: final wide shot as he lowers the can, camera completes orbit to original angle, showcasing the kitchen space, lighting, and dynamic movement)."
-}
+}