PlagueKind - Minimax H3 - LTX2.3 Workflow - Ease of use - Eros | Sulphur Compatible | FaceID
## Minimax H3
## V5.6 Notes
- Unified resize now has Aspect ratio selector
- Changed default sampler to sa_solver
- I needed to make this because changing unified resize breaks old workflows. update nodepack before loading the workflow
## V5.5 Notes
- Major'ish Overhaul of the layout, and wiring.
- Disabling Sparse falls back to comfy attention.
Sparse Density Notes:
0.80-0.85 - "match pytorch quality"
0.9 - Default - Minimal Degredation 15-20% faster than 0.85
0.95 - HUGE boost for longer hi res generations, can introduce minor degredation
## V5 Notes
- I've added my Sparse attention node. No other speedups required. up to 2.5x speedup.
- Switched to dareties lora. Use with shift 8/3 0.8 strength 6-8 steps. this lora works great for all modes.
- Memory Chunking if you go to high with sparse, it causes slow down. for me beyond 2-4 caused slowdown on 16gb vram.
- Don't mess with the sparse settings on the subgraph unless you have to.
- Make sure to update my node pack to atleast 1.3.1
- Make sure you're on the newest pytorch+cu combo supported.
### Support
If you find this project useful and want to support development:
Monero (XMR):
865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7p## V4 Notes
- Sage begone, welcome comfy attn. update comfyui, sage without sage issues.
- Switched to 4 step 768 lora. Use with shift 6/3 0.8 strength 6-8 steps, or 1.0 with 4-6 steps.
- Added hybrid model so you don't need to swap anymore, start with 25. 30 higher quality slightly less ref strength.
- When ref swapping a video use er_sde/beta
- Added original audio mode, avoids degradation and vae decode / Allows you to make music videos etc. it's in the reference subgraph.
- Added minimax mode and mode selection to lora stack, so update node pack.
## V3.1 Notes
- Sage begone, welcome comfy attn. update comfyui, sage without sage issues.
- Switched to 8 step new lora, these 2 new things combined behave very close to non turbo non sage.
## V3 Notes
- Added R2V mode.
## V2.1 Notes
- Added preview node.
## V2 Notes
- Added int8 convrot video vae
- Kijai resized turbo lora, best at 0.6-0.7 strength.
- Memory chunks return. huge speed up for me. 4 was pretty much the cap for 16gb vram.
---
## Requirements
- [ComfyUI-PlagueKind-Nodes](https://github.com/PlagueKind/Comfyui-PlagueKind-Nodes)
- [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
- [Rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
- [Nvidia_RTX_Nodes_ComfyUI](https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI)
- [ComfyUI-VideoHelperSuite](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
- [ComfyUI-H3-Motion-Context-MultiRef
](https://github.com/seitanism/ComfyUI-H3-Motion-Context-MultiRef)
---
### Diffusion Model
[minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors)
[minimax_h3_hybrid_fl2va_ref2va_b30-49-int8.safetensors](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b30-49-int8.safetensors)
### VAE
[minimax_h3_audio_vae_fp32.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors)
[minimax_h3_video_vae_int8_convrot.safetensors](https://huggingface.co/Kijai/MiniMax-H3-experimental/resolve/main/minimax_h3_video_vae_int8_convrot.safetensors)
[vae_approx/taeh3.safetensors](https://huggingface.co/Kijai/MiniMax-H3-TAE/resolve/main/vae_approx/taeh3.safetensors)
### Text Encoder
[qwen3vl_32b_minimax_h3_int8_convrot.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors)
### Frame Interpolation
[film_net_fp16.safetensors](https://huggingface.co/Comfy-Org/frame_interpolation/resolve/main/frame_interpolation/film_net_fp16.safetensors)
### Turbo Lora
[dareties_v4_step600_comfy_fro.safetensors](https://huggingface.co/silveroxides/MiniMax-H3_tests/resolve/main/minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600_comfy_fro.safetensors)
supporting file required.
make the folder h3_adaln in models.
[h3_silu_temb_grid.safetensors](https://huggingface.co/deAPI-ai/minimax-h3-33b-int8/resolve/main/loras/h3_silu_temb_grid.safetensors)
---
## V1.5 Notes
- Added the option for T2V.
- When using T2V, set scale mode of Target Dimension node to Dimensions (WxH) or you will get a square video.
- Added FSR RCAS sharpening, 0.3 strength looks natural.
- FFLF is unreliable at res above 640, don't know why, it's like this on default WF as well.
- Target Dimension node controls the generation size only, refs are passed elsewhere.
- Don't touch the other hidden resize nodes as they only control ref size and are set to max.
- Euler + Linear Quadratic seems to be the best combo with between 15-20 steps.
- I didn't add easy cache because it ruins the quality noticeably,
- If you have time or a powerful card, try heunpp2 sampler. It's phenomenal.
## LTX 2.3
## V8 Notes
- Updated everything to LTX 2.5.
- Lora stack still seems to work.
## V7 Notes
- Added FaceID tracking.
- Changed DMD Lora for smoother output.
- Steps required 4-10. 6 seems good enough.
- Pruna Video VAE for speed boost.
- Int8 Convrot TE.
- Any FPS choice seems to work.
- Increase IMG Preprocess for extra movement.
- ComfyUI broke subgraphs so I adapted.
- Make sure you're on atleast 10s-nodes 1.9.6
### Support
If you find this project useful and want to support development:
Monero (XMR):
`
865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7p
`
Thank you for your support.
## V6 Notes
- Updated Eros to 1.4 much better.
- You can disable all optional lora apart from distill and eros if you want. The others do enhance the output though.
- Added Volume Boost option.
- Using DMD for audio and video.
- Use 0.715, 0.2412, 0.0 2x FPS sigma if you see jitter
## V5 Notes
- Using decoupled distillation for video and audio, and smaller loras. - REQUIRED.
- Updated to Eros 1.3
- FPS bumped to 30fps default.
- 60fps native with 120fps via 2x now compatible.
- Added an actually useable NVFP4 model, even though int8 convrot is more or less the same size and speed.
- You can regain speed by disabling one of the optional loras, start with sulphur.
## V4 Notes
- New reference node injects reference directly into the model, mostly eliminating any drift.
- New DMD distill lora, fixes issues of original distill which broke decompression method and prompt adherence, no need to write a paragraph anymore. Preprocess returns for enhanced movement.
- Use 6-13 steps, default 8-9.
- INT8 Triton VAE for fast decode.
- INT8 Convrot added for increased speed on all cards.
- JoyAI lora another lora similar to omninft, this helps adherence, dialog, interaction and more.
- If using the nvfp4 transformer make sure to describe anatomy that isn't already visible in the image. I don't recommend using the nvfp4 text encoder for realism.
---
## V3 Notes
TURN DOWN NEW EROS LORA TO 0.7 I SET IT AT 0.9 and it doesn't look correct like that.
- Added Prompt Relay Node for advanced timeline prompting while keeping simplicity.
- Added VHS save video. AV1/WEBM CRF0 produces virtually lossless video with extremely small file size.
- Updated Eros Lora.
- Moved RTX VSR into sampler graph.
## V2 Notes
- Upscale and 2x fps have added tiled sampler to improve speed. if you don't like the output, experiment with cfg pp sampler, or add 0.715 to sigma. Current sigma are done on purpose to retain identity and quality of first pass.
## V1 Notes
- Focused Mainly on Speed and ease of usability while retaining single pass quality and likeness.
- First pass is more or less optimized.
- Identity Retention is pretty good.
- First pass balanced for speed and quality.
- Upscale and 2x fps work but will be optimized later.
- I would avoid using RTX VSR with the whole chain active unless you have like 96gb ram.
- Nvfp4 models are only for Blackwell, using them will not save you speed if you don't have Blackwell, but they should still work.
- Default settings of the first pass should produce good quality and sharp. I personally like to use this for a single pass and just render straight at 1280.
- Bumped FPS to 30 because it looks way bettwen imo, quality seems clearer overall and it gets rid of that ltx base fps look.
---
## Memory Chunk Note
- This saves vram and speeds up it/s especially on longer videos
- With 16gb of VRAM 3 chunks seems optimal for me
- If you get triton errors, enter the model loader subgraph and disable use triton kernels
---
## Requirements
- [ComfyUI-PlagueKind-Nodes](https://github.com/PlagueKind/Comfyui-PlagueKind-Nodes)
- [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
- [Rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
- [10s-nodes](https://github.com/TenStrip/10S-Comfy-nodes)
- [Nvidia_RTX_Nodes_ComfyUI](https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI)
---
### Model Choices
[ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors)
### Latent Upscalers
[ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors)
[ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/latent_upscale_models/ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors)
### VAE
[ltx-2.5-audio-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/vae/ltx-2.5-audio-vae-bf16.safetensors)
[ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors)
### Text Encoder
[gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors](https://huggingface.co/Lightricks/LTX-2.5/resolve/main/text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors)
### FaceID LoRA
[Best_FaceID_v1.0_LoRA.safetensors](https://huggingface.co/Alissonerdx/LTX-Best-Face-ID/resolve/main/Best_FaceID_v1.0_LoRA.safetensors)
### Optional LoRA
[LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors](https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/loras/LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors?download=true)
[LTX_SulphurEXP_LoRA_fro99-avgrank105.safetensors](https://huggingface.co/maximsobolev275/LTX-SulphurExperimental-LoRA-Optimized/resolve/main/LTX_SulphurEXP_LoRA_fro99-avgrank105.safetensors)
[JoyAI-Echo-content_r256.safetensors](https://huggingface.co/TenStrip/LTX2.3_JoyAI_Lora_Extracted/resolve/main/JoyAI-Echo-content_r256.safetensors)
[LTX_10Eros-v15_LoRA_fro99-avgrank103.safetensors](https://huggingface.co/maximsobolev275/LTX-10Eros-LoRA-r768/resolve/main/LTX_10Eros-v15_LoRA_fro99-avgrank103.safetensors)
### Optional Model converted by me
[10Eros_v1_experimental_nvfp4.safetensors](https://huggingface.co/Plaguekind/10Eros_v1_experimental_nvfp4/resolve/main/10Eros_v1_experimental_nvfp4.safetensors?download=true)
Description
## V6 Notes
- Updated Eros to 1.4 much better.
- You can disable all optional lora apart from distill and eros if you want. The others do enhance the output though.
- Added Volume Boost option.
- Using DMD for audio and video.
- Use 0.715, 0.2412, 0.0 2x FPS sigma if you see jitter
FAQ
Comments (30)
OMG This is the best workflow I've ever tried absolutely insane quality with the newest eros model!!!!
Thanks alot, and thanks for the tip. I'm glad you found it useful
For some reason, I'm unable to use SageAttention, which the workflow doesn't detect - even though it is definitely installed and works fine with the exact same node in other workflows. This was also the case with versions 5 and 4.
Other than that, everything works great, thanks 👍
Make sure you have everything up to date. Never heard of an issue like this.
No matter what I do, I can't get anything except for a black screen but I can hear audio. I tried the nvfp4 and v1.4 models from Civit, enabling and disabling triton VAE, and I never turned on sage attention
Edit: Nevermind, they were being generated but showing up as black in the comfyui preview. In file explorer I could see them
Not normal at all. Make sure everything is up to date.
can this run Sulphur 2?
Yeah sure can. It's for i2v though not t2v, so eros makes more sense.
@Plaguekind i've just tried Sulphur distill fp8 with this workflow, i gotta give it to you thanks to you i finally can run the model i want + the prompt following is almost perfect, what i did was i slapped Sulphur Distill fp8 as the main model, and only used one Eros lora of the lowest size, so far it's running perfectly and the speed is okayish, my only complaint is that the sound is not perfect and doesn't follow the prompt when it comes to saying things/grunting etc.
but the slapping sound is working so that's fine by me.
i tried all the other models to make the girl for example motionless and expressionless like a doll, none did that except for the model i am using now, i'm trying to make it run a bit faster but dunno if any lora could help me do that without ruining the quality.
also i found out that the spatial and temporal upscalers kill the prompt adherence for some reason ????
without them i get much more solid prompt adherence + video quality, it's weird i know but that's how it is from my testings so far.
thank you SO much for providing this workflow, so far so good !
let me know if Eros can do the prompt i mentioned above if you had the time, i will switch to it and try it out if it does, cuz i've been trying tons of models as i mentioned and none did it better than the one i mentioned.
@Plaguekind i just wish the convrot version of Eros can give me the same results and prompt following, last time i tried, it generates random moaning and facial expressions, i used this one:
https://huggingface.co/Winnougan/10Eros-INT8-Convrot
and it follows everything except for facial expressions, any ideas? convrot would be much faster for my case.
@Plaguekind oh i forgot to mention that i am using it for i2v only.
@necroryona my personal reccomendation is to use it how i have it setup. Use the convrot dev transformer, and the sulphur + eros1.4 loras.
@necroryona the reason for the okish speed is because it's not running half res like other wf. It runs at the res you set. This workflow should be faster than others if you have everything enabled.
@Plaguekind i'll def try to follow the models you mentioned in the WF and give it another try tomorrow, thank you again !
Hi, I love your workflow. Could you please tell me how I could connect to your current workflow the LTX Face Identity Reinforcer from 10S nodes? I get very quick generations with good quality.
Just replace the connections of the reference guide node with that one, works fine. And thank you!
@Plaguekind I asked ChatGPT to help me modify the workflow. It added the Best-Face-ID LoRA before the identity path, kept your existing LTXReferenceConditioning, then connected LTXFaceIdentityReinforcer after it using the video VAE, the preprocessed reference image, the same image as reference_image_2, and the target_latent from LTXVImgToVideoInplaceKJ. The Reinforcer model output is then routed to the sampler.
Do you think this chained setup is okay, or would you recommend fully replacing LTXReferenceConditioning with LTXFaceIdentityReinforcer instead?
@john455454674 you need the kj node for it to work. You can literally just replace the piping of the reference node and replace it with the faceidnode, and you can safely put the faceid lora in the stack. Im working on a v7 that uses the node. But I'm still deciding if i like it more
@Plaguekind, thanks; I did it, but I will wait and try your v7 also.
How to add preview during generation? I don't see
You'll have to pipe it in. I havent gotten around to that yet.
https://civitai.red/models/2788491/ltx-23-turbo-distilled-v11-int8-row-convrot-hq?modelVersionId=3142108
This thing is bloody amazing. Much better results with no distilled/DMD, faster and just amazing clarity, motion, prompt adherence and facial preservation. Reminds me more of prompting WAN.
Check it out. I know you're looking to improve your workflow.
It's a row wise pre distilled. I wouldn't benefit from row wise. But i will test it anyway. Thanks
@Plaguekind Right, I'm just fumbling in the dark here. There's a dev version too.
You do what you do, it's your workflow :) Looking forward to v7.
I gave this a go and was pretty blown away
@lonlon1990 I'll test it today
Thanks for the workflow. Is there a solution to get more motion? Characters move very little when I generate a video. Even if I describe something like "Wild, extreme motion" , they only move a little.
Any tips would be helpful.
I’ve been trying different workflows all week. Even tried making my own from scratch. I’m happy I found this one. It’s easy to understand and it’s not a mess! I recommend adding a picture of the workflow to the preview images of this post. Other people should see just how simple you made this workflow
It's already there mate.
possible to add ltx Director?