studio_ghibli_wan14b_t2v_v01.safetensors

293 MB · SafeTensor

studio_ghibli_wan14b_t2v_v01_dataset.zip

286 MB · Other

Requiere membresía

<ghibli-style-wan>

This LoRA is on OpenMuse, a curated initiative dedicated to open-source video LoRAs and the creative works they enable. Focused on models like Wan2.1, LTX-Video, and HunyuanVideo, OpenMuse highlights high-quality tools and artwork from across the ecosystem. Rooted in the community, OpenMuse is a growing home for open, collaborative AI art, designed to inspire creators, spark curiosity, and offer something you'd feel proud to share, even with someone skeptical of AI-generated art. Description I am very happy to share my magnum opus LoRA, which I've been working on for the past month since Wan came out. This is indeed the best LoRA on Civitai I have ever trained, and I have to say (once again) - is an amazing model. LoRA was trained for ~90 hours on an RTX 3090 with using a mixed dataset of 240 clips and 120 images. This could have been done faster, but I was obsessed with pushing the limits to create a state-of-the-art style model. It’s up to you to judge if I succeeded. Usage The trigger phrase is Studio Ghibli style - all captions for training data were prefixed with these words. All clips I publish in gallery are raw outputs using a LoRA with a base Wan-T2V-14B model (although latest videos also may include self-forcing LoRA for inference acceleration, read more below), without further post-processing, upscaling, or interpolation. Compatibility with other LoRAs and with Wan-I2V models has not been tested. Workflows are embedded with each video (so you can just download and drag video into ComfyUI to open it). As an example, for workflow (based on ), which uses the (created by ), extracted from lightx2v's . I chose version made by blyss (and not the ), because, from my tests, it offers maximum compatibility and only accelerates inference, without any additional detailing or stylistic bias. (This is also the reason why I stick to the base Wan model and do not use merges like or .) I'm using the acceleration LoRA with the UniPC sampler (and occasionally DPM++). In my experience, UniPC performs better for 2D animation than LCM, which tends to lean more toward realism, which I want to avoid. Usually I also apply NAG node, so I can use negative prompts with CFG=1. From initial testing, compared to the older workflow with TeaCache, aside from the huge speed gain (a 640×480×81 6-step clip renders in ~1 minute instead of 6 on an RTX 3090), it also slightly improves motion smoothness and text rendering. The are also very impressive in terms of speed and quality preservation. I'm using a , but the 32 and 64 versions also produce great results. Here's an example of the workflow . I found out lowering lightx2v LoRA strength to 0.9, increasing number of steps to 8, and using either UniPC or DPMPP scheduler gives pretty good outputs. The obvious downside is that output usually leans to default Wan's "realistic 3D" style. Counter this by increasing number of steps, lowering strength of acceleration LoRA and increasing strength of style LoRA. Also you can try to replace lightx2v LoRA with , it might occasionally give slightly better motion. And is "legacy" workflow in JSON format. It was used to generate 90% videos in gallery for this LoRA. It was also build on wrapper nodes and included a lot of optimizations (more information ), including fp8_e5m2 checkpoints + torch.compile, SageAttention 2, TeaCache, Enhance-A-Video, Fp16_fast, SLG, and (sometimes) Zero-Star (some of these migrated to new workflow as well), but rendering a 640x480x81 clip still took about 5 minutes (RTX 3090) in older workflow. Although the legacy workflow demonstrates slightly superior quality in a few specific areas (palette, smoothness), the 5x slowdown is a significant and decisive drawback, being the reason I migrated to lightx2v-powered version. Prompting To generate most prompts, I usually apply the following meta-prompt in ChatGPT (or Claude, or any other capable LLM), that helps to enhance "raw" descriptions. This prompt is based on official by Wan developers and looks like this: You are a prompt engineer, specializing in refining user inputs into high-quality prompts for video generation in the distinct Studio Ghibli style. You ensure that the output aligns with the original intent while enriching details for visual and motion clarity. Task Requirements: - If the user input is too brief, expand it with reasonable details to create a more vivid and complete scene without altering the core meaning. - Emphasize key features such as characters' appearances, expressions, clothing, postures, and spatial relationships. - Always maintain the Studio Ghibli visual aesthetic - soft watercolor-like backgrounds, expressive yet simple character designs, and a warm, nostalgic atmosphere. - Enhance descriptions of motion and camera movements for natural animation flow. Include gentle, organic movements that match Ghibli's storytelling style. - Preserve original text in quotes or titles while ensuring the prompt is clear, immersive, and 80-100 words long. - All prompts must begin with "Studio Ghibli style." No other art styles should be used. Example Revised Prompts: "Studio Ghibli style. A young girl with short brown hair and curious eyes stands on a sunlit grassy hill, wind gently rustling her simple white dress. She watches a group of birds soar across the golden sky, her bare feet sinking slightly into the soft earth. The scene is bathed in warm, nostalgic light, with lush trees swaying in the distance. A gentle breeze carries the sounds of nature. Medium shot, slightly low angle, with a slow cinematic pan capturing the serene movement." "Studio Ghibli style. A small village at sunset, lanterns glowing softly under the eaves of wooden houses. A young boy in a blue yukata runs down a narrow stone path, his sandals tapping against the ground as he chases a firefly. His excited expression reflects in the shimmering river beside him. The atmosphere is rich with warm oranges and cool blues, evoking a peaceful summer evening

SHA256

DD2FE1258DB4288133F866103638F8A635CEE065E50AEF818B0F0569335E9EAB

ID de modelo de ModelTrends.ai

#11918

Tamaño

576x576

Tamaño

624x624

Tamaño

640x480

Tamaño

640x480

Tamaño

480x832

Tamaño

544x544