PlagueKind - LTX2.3 Workflow - I2V - NVFP4 | INT8 Convrrot | Ease of use - Eros | Sulphur Compatible
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
Sage Attention (optional)
Model Choices
ltx-2.3-22b-dev-nvfp4.safetensors
ltx-2.3-22b-dev_transformer_only_int8_convrot.safetensors
Latent Upscalers
ltx-2.3-spatial-upscaler-x2-1.1.safetensors
ltx-2.3-temporal-upscaler-x2-1.0.safetensors
VAE
LTX23_audio_vae_bf16.safetensors
LTX23_video_vae_bf16.safetensors
Dual CLIP (Encoder and Projection) Choices
gemma_3_12B_it_fp4_mixed.safetensors
gemma_3_12B_it_fp8_scaled.safetensors
ltx-2.3_text_projection_fp8.safetensors
ltx-2.3_text_projection_bf16.safetensors
Distilled LoRA
LTX2.3_DMD_reshaped_r256.safetensors
Optional LoRA
LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors
LTX_10Eros-v12_LoRA_fro99-avgrank91.safetensors
JoyAI-Echo-content_r256.safetensors
LTX_SulphurEXP_LoRA_fro99-avgrank105.safetensors