Natalie Hershlag, known professionally as Natalie Portman, is an Israeli-born American actress. Born: 9 June 1981.
Fuel my GPU's caffeine addiction! (Ko-fi)
https://ko-fi.com/cyberaimania
I would really appreciate it if you could rate and comment on my lores. Your feedback, whether positive or constructive criticism, truly gives me a "kick" and the energy to create new projects. Thank you in advance for your engagement!
Trained on the Wan2.1 T2V 14B base model using diffusion-pipe.
*Dataset:** 112 photos.
*Resolution:** Trained targeting 512px area (aspect ratio bucketing enabled).
*Key Settings:** Rank 32 / LR 3e-5 / 10 Epochs / 10 Repeats (~11.2k total steps).
*Optimizations Used:** Fast settings with Grad Accum = 1 / Blocks Swap = 0, AdamW8bitKahan optimizer, active FP8 precision for DiT & LLM, and Unsloth activation checkpointing.
Trigger: pornatx
All my video samples were generated using EPOCH 09, and this same EPOCH 09 is available for download from this site.
If you think Natalie looks better when generated with a different EPOCH, feel free to check out:
https://huggingface.co/ArkaDio81/portman/tree/main
I made a simple configurator for TOML files. This is an ALPHA version — it only suggests settings, so you need to be cautious and keep that in mind. It's mainly intended for GPUs with 24GB of VRAM.
https://drive.google.com/file/d/1Kr-zRImaRJ0by6wO2sQiGnxBWl2jCM_S/view?usp=sharing
Check out my other LoRAs:
https://civarchive.com/user/CyberAImania
Enjoy!
Description
FAQ
Comments (14)
Looks awesome. Do you use an LLM to help come up with your prompts?
Thank you.
Yes, I use prompts with the help of Gemini on the free https://aistudio.google.com/.
@CyberAImania Do you go with the official Wan instructions or do you have a custom one you like?
@b1996354 What do yuu mean ? custom one what ?
Hello CyberAlmania,
I really like your work, and thank you so much for sharing all the detailed data. I'm already looking forward to your next creations!
I have a question. With WAN, I believe the issue is that the models we train in the diffusion pipe are quite different from the ones we end up using for generation.
Could you please tell me exactly which models you used for training and which ones for generation (a link would be very helpful if that's not a problem), since you've achieved such excellent results?
Thank you in advance!
Hello there!
Thank you so much for the incredibly kind words about my work and for appreciating the shared data! That really means a lot to me, and I'm excited about future creations too!
That's a very good question you're asking about the models used for training vs. generation in Diffusion-Pipe for Wan. Honestly, I'm definitely not an expert yet – I'm very much still learning and experimenting myself! But I'm happy to share what I've figured out so far, because that's what this awesome open-source community is all about.
So, for all my LoRA training runs in Diffusion-Pipe, I've been using the Wan2.1 T2V 14B base model. More specifically, the main transformer weights I'm using are the FP8 version. The exact file is called:
Wan2_1-T2V-14B_fp8_e5m2.safetensors
I downloaded it from the repositories linked by the Diffusion-Pipe creator. You can usually find this specific file and other necessary base components (like the VAE and T5) within the Wan-AI and Comfy-Org repositories on Hugging Face.
Here are the links to the main locations where I get the base model components:
Main Wan2.1 Model Hub: https://huggingface.co/Wan-AI
ComfyUI Repackaged Models (often includes VAE, FP8 DiT): https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged
You'll need a few files from these places – the main model structure, the transformer weights (I use that specific FP8 one), the VAE, and the T5. The Diffusion-Pipe documentation usually clarifies which files go where.
Now, regarding generation – that's where things get interesting! While I train my LoRAs on that specific T2V 14B FP8 base model within Diffusion-Pipe, for generating videos (like the ones you see) I actually use the gguf Q_8 version of the Wan2.1 model. This is a quantized version optimized for faster inference, and I use it with a different generation interface (specifically, one that supports GGUF models).
https://huggingface.co/city96/Wan2.1-T2V-14B-gguf/resolve/main/wan2.1-t2v-14b-Q8_0.gguf
So there's a bit of a mix-and-match there – training the LoRA on a specific base model/format, and then applying that LoRA when generating with a different, inference-optimized version of the base model (the GGUF). It seems to work quite well!
I hope this explanation helps clear things up a bit! I'm still experimenting with different settings and approaches myself, and it's awesome to learn from and share with others in the community.
Thanks again for your kind comments!
Can you post an image or video containing your workflow? I'm new to WAN T2V and am not getting anything close to the same quality results as your sample videos with a workflow I found on Civit.
I'm using all of his workflow
https://civitai.com/models/1309324/txt-to-video-simple-workflow-wan21-or-gguf-or-lora-or-upscale-or-teacache
@CyberAImania Okay. Just had some time to look at that workflow. There different workflows in packed into that link. Are you using the pure T2V version or the Flux -> I2V version?
@Clocksmith FLUX ??, you asking about image FLUX or Wan Viedeo ?
I using wan2.1-t2v-14b-Q8_0.gguf for all my generations
@CyberAImania How long does it take and what hardware , VRAM are you using?
@svenhimmelvarg RTX 4090 24GB VRAM, AMD 7950X, 64GB RAM DDR5. Between 5 and 24h. Depends how big dataset and training resolution
@CyberAImania There is a Flux first frame then I2V workflow in the link you sent. Just wanted to make sure you weren't using that. Subsequent tests on my side verify this is indeed a full WAN T2V lora, not a strange Flux / I2V lora, as I was first concerned about when you sent your link.
@Clocksmith Sorry for the confusion. I mainly wanted to say that the other author has very good workflows.
https://civitai.com/user/UmeAiRT/models?sort=Highest+Rated&baseModels=Wan+Video
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.