"""Sample 16-frame I2V clips with Anima Turbo + the motion module through the ComfyUI API.""" import json, sys, time, urllib.request from pathlib import Path sys.path.insert(0, str(Path(__file__).parent)) from gen_h3 import API, DATA, post, upload # noqa: E402 CASES = [("c0001", "var1", "seen"), ("c0006", "var2", "seen"), ("c0040", "var0", "new"), ("c0057", "var1", "new")] MOTION = {"c0001": "waves his hand and laughs", "c0006": "raises one arm and stretches", "c0040": "turns her head toward the camera and smiles", "c0057": "waves her hand and laughs"} def graph(image, prompt, motion, prefix, ref_image=None, mask_grow=None, denoise=1.0, mask_query="character", invert=False): model = ["1", 0] g = { "1": {"class_type": "UNETLoader", "inputs": {"unet_name": "anima/anima-turbo-v1.0-int8-convrot.safetensors", "weight_dtype": "default"}}, "2": {"class_type": "CLIPLoader", "inputs": {"clip_name": "qwen_3_06b_base.safetensors", "type": "stable_diffusion", "device": "default"}}, "3": {"class_type": "VAELoader", "inputs": {"vae_name": "qwen_image_vae.safetensors"}}, "4": {"class_type": "CLIPTextEncode", "inputs": {"clip": ["2", 0], "text": prompt}}, "5": {"class_type": "CLIPTextEncode", "inputs": {"clip": ["2", 0], "text": "worst quality, low quality, blurry, text, watermark"}}, "6": {"class_type": "LoadImage", "inputs": {"image": image}}, "7": {"class_type": "ImageScale", "inputs": {"image": ["6", 0], "upscale_method": "lanczos", "width": 384, "height": 576, "crop": "center"}}, "8": {"class_type": "CosmosPredict2ImageToVideoLatent", "inputs": {"vae": ["3", 0], "width": 384, "height": 576, "length": 61, "batch_size": 1, "start_image": ["7", 0]}}, "20": {"class_type": "KSampler", "inputs": {"positive": ["4", 0], "negative": ["5", 0], "latent_image": ["8", 0], "seed": 3, "steps": 8, "cfg": 1.0, "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}}, "24": {"class_type": "LatentCutToBatch", "inputs": {"samples": ["20", 0], "dim": "t", "slice_size": 1}}, "21": {"class_type": "VAEDecode", "inputs": {"vae": ["3", 0], "samples": ["24", 0]}}, "22": {"class_type": "CreateVideo", "inputs": {"images": ["21", 0], "fps": 8.0}}, "23": {"class_type": "SaveVideo", "inputs": {"video": ["22", 0], "filename_prefix": prefix, "format": "mp4", "codec": "h264"}}, } if ref_image: # R2V: character sheet through Anima Ref Adapter, no start image g["31"] = {"class_type": "LoadImage", "inputs": {"image": ref_image}} g["32"] = {"class_type": "AnimaRefAdapter", "inputs": {"model": model, "vae": ["3", 0], "image": ["31", 0], "lora_name": "anima_ref_adapter_v2_r64.safetensors", "ref_guidance": 1.0, "correction": "none"}} model = ["32", 0] del g["8"]["inputs"]["start_image"] if mask_grow is not None: # background frozen: keyframe latent in every frame, only the WeVisOne character matte is denoised g["40"] = {"class_type": "VAEEncode", "inputs": {"pixels": ["7", 0], "vae": ["3", 0]}} for i in range(41, 45): g[str(i)] = {"class_type": "LatentConcat", "inputs": {"samples1": [str(i - 1), 0], "samples2": [str(i - 1), 0], "dim": "t"}} g["45"] = {"class_type": "WeVisOneVLLoader", "inputs": {"clip_name": "wevisone_vl_4b_int8_convrot.safetensors"}} g["46"] = {"class_type": "WeVisOneVLPerceive", "inputs": {"clip": ["45", 0], "image": ["7", 0], "task": "matting", "query": mask_query, "max_length": 1024, "individual_masks": False}} g["47"] = {"class_type": "ThresholdMask", "inputs": {"mask": ["46", 3], "value": 0.3}} g["48"] = {"class_type": "GrowMask", "inputs": {"mask": ["47", 0], "expand": mask_grow, "tapered_corners": True}} mask = ["48", 0] if invert: # redraw everything but the (grown) matte g["52"] = {"class_type": "InvertMask", "inputs": {"mask": mask}} mask = ["52", 0] g["49"] = {"class_type": "SetLatentNoiseMask", "inputs": {"samples": ["44", 0], "mask": mask}} g["50"] = {"class_type": "MaskToImage", "inputs": {"mask": mask}} g["51"] = {"class_type": "SaveImage", "inputs": {"images": ["50", 0], "filename_prefix": prefix + "_mask"}} g["20"]["inputs"]["latent_image"] = ["49", 0] g["20"]["inputs"]["denoise"] = denoise g["20"]["inputs"]["denoise"] = denoise if motion: g["30"] = {"class_type": "AnimaMotionModule", "inputs": {"model": model, "motion_name": motion}} model = ["30", 0] g["20"]["inputs"]["model"] = model return g def main(motion, tag, r2v=False, mask_grow=None, denoise=1.0, mask_query="character", invert=False): for char, var, kind in CASES: meta = json.loads((DATA / char / "meta.json").read_text()) prompt = f"masterpiece, best quality, score_7, safe, {meta['character_tags']}, {MOTION[char]}, anime screencap" if not meta["girl"]: prompt = prompt.replace("her ", "his ") image = upload(DATA / char / f"{var}.png") ref_image = upload(DATA / char / "ref.png") if r2v else None prefix = f"anima_motion_eval/{tag}_{kind}_{char}" pid = post("/prompt", json.dumps({"prompt": graph(image, prompt, motion, prefix, ref_image, mask_grow, denoise, mask_query, invert)}).encode(), {"Content-Type": "application/json"})["prompt_id"] t = time.time() while pid not in (h := json.load(urllib.request.urlopen(f"{API}/history/{pid}"))): time.sleep(2) s = h[pid]["status"] print(prefix, s["status_str"], f"{time.time() - t:.0f}s", flush=True) if s["status_str"] != "success": print(json.dumps(s["messages"])[-1200:]) if __name__ == "__main__": grow = next((int(a[5:]) for a in sys.argv[3:] if a.startswith("mask=")), None) denoise = next((float(a[8:]) for a in sys.argv[3:] if a.startswith("denoise=")), 1.0) query = next((a[6:] for a in sys.argv[3:] if a.startswith("query=")), "character") main(sys.argv[1] if sys.argv[1] != "none" else None, sys.argv[2], "r2v" in sys.argv[3:], grow, denoise, query, "invert" in sys.argv[3:])