Upload folder using huggingface_hub

README.md

@@ -0,0 +1,115 @@
+---
+license: apache-2.0
+language:
+- en
+- zh
+tags:
+- image-to-video
+- lora
+- replicate
+- text-to-video
+- video
+- video-generation
+base_model: "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+pipeline_tag: text-to-video
+# widget:
+#   - text: >-
+#       prompt
+#     output:
+#       url: https://...
+instance_prompt: nivedan
+---
+
+# Wan2.1 Lora
+
+<Gallery />
+
+## About this LoRA
+
+This is a [LoRA](https://replicate.com/docs/guides/working-with-loras) for the Wan2.1 14b video generation model.
+
+It can be used with diffusers or ComfyUI, and can be loaded against both the text-to-video and image-to-video Wan2.1 models.
+
+It was trained on [Replicate](https://replicate.com/) using AI toolkit: https://replicate.com/ostris/wan-lora-trainer/train
+
+
+## Trigger words
+
+You should use `nivedan` to trigger the video generation.
+
+
+## Use this LoRA
+
+Replicate has a collection of Wan2.1 models that are optimised for speed and cost. They can also be used with this LoRA:
+
+- https://replicate.com/collections/wan-video
+- https://replicate.com/fofr/wan2.1-with-lora
+
+### Run this LoRA with an API using Replicate
+
+```py
+import replicate
+
+input = {
+    "prompt": "nivedan",
+    "lora_url": "https://huggingface.co/NIVEDAN/wan2.1-lora/resolve/main/wan2.1-14b-nivedan-lora.safetensors"
+}
+
+output = replicate.run(
+    "fofr/wan2.1-with-lora:f83b84064136a38415a3aff66c326f94c66859b8ad7a2cb432e2822774f07b08",
+    model="14b",
+    input=input
+)
+for index, item in enumerate(output):
+    with open(f"output_{index}.mp4", "wb") as file:
+        file.write(item.read())
+```
+
+### Using with Diffusers
+
+```py
+pip install git+https://github.com/huggingface/diffusers.git
+```
+
+```py
+import torch
+from diffusers.utils import export_to_video
+from diffusers import AutoencoderKLWan, WanPipeline
+from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
+
+model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
+flow_shift = 3.0  # 5.0 for 720P, 3.0 for 480P
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift)
+pipe.to("cuda")
+
+pipe.load_lora_weights("NIVEDAN/wan2.1-lora")
+
+pipe.enable_model_cpu_offload() #for low-vram environments
+
+prompt = "nivedan"
+negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+
+output = pipe(
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    height=480,
+    width=832,
+    num_frames=81,
+    guidance_scale=5.0,
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+```
+
+
+## Training details
+
+- Steps: 2201
+- Learning rate: 0.0001
+- LoRA rank: 40
+
+
+## Contribute your own examples
+
+You can use the [community tab](https://huggingface.co/NIVEDAN/wan2.1-lora/discussions) to add videos that show off what you’ve made with this LoRA.

config.yaml

@@ -0,0 +1,62 @@
+job: custom_job
+config:
+  name: wan_train_replicate
+  process:
+  - type: custom_sd_trainer
+    training_folder: output
+    device: cuda:0
+    trigger_word: nivedan
+    network:
+      type: lora
+      linear: 40
+      linear_alpha: 40
+    save:
+      dtype: float16
+      save_every: 2202
+      max_step_saves_to_keep: 1
+    datasets:
+    - folder_path: input_images
+      caption_ext: txt
+      caption_dropout_rate: 0.05
+      shuffle_tokens: false
+      cache_latents_to_disk: false
+      cache_latents: true
+      resolution:
+      - 632
+    train:
+      batch_size: 1
+      steps: 2201
+      gradient_accumulation_steps: 1
+      train_unet: true
+      train_text_encoder: false
+      gradient_checkpointing: false
+      noise_scheduler: flowmatch
+      timestep_type: sigmoid
+      optimizer: adamw8bit
+      optimizer_params:
+        weight_decay: 0.0001
+      lr: 0.0001
+      ema_config:
+        use_ema: true
+        ema_decay: 0.99
+      dtype: bf16
+    model:
+      name_or_path: Wan-AI/Wan2.1-T2V-14B-Diffusers
+      quantize: false
+      arch: wan21
+    sample:
+      sampler: flowmatch
+      sample_every: 2202
+      width: 832
+      height: 480
+      num_frames: 33
+      fps: 16
+      prompts: []
+      neg: ''
+      seed: 42
+      walk_seed: true
+      guidance_scale: 5
+      sample_steps: 30
+meta:
+  name: wan_train_replicate
+  version: '1.0'

wan2.1-14b-nivedan-lora.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e019a709a1e69ba9792054de85c8d7f557a8ea57364558d2dcf3cf3c856f5b41
+size 383484864