base_model: Lightricks/LTX-2.5 base_model_relation: adapter license: other license_name: ltx-2.x-community-license-agreement license_link: https://huggingface.co/lvladikov/LTX-2.5-Stage1-Distill-2step-LoRA/blob/main/LICENSE-2_x library_name: diffusers tags:

lora
text-to-video
image-to-video
distillation
step-distillation
ltx-2.5
comfyui pipeline_tag: text-to-video widget:
text: >- A young woman with freckles and windswept auburn hair stands by a tall window in soft daylight. A breeze blows strands of hair across her face; she tucks them behind her ear, turns her head from the window to look straight into the camera and smiles softly, then laughs, her eyes still on the camera to the last frame. Shallow depth of field, photographic realism, sharp detail. Sound: a gentle breeze and distant street noise. output: url: assets/videos/t2v_sample05_1152x640.mp4
text: >- Live-action documentary footage shot on 35 mm film: a real young inventor, not animated, not 3D, not CGI, with oversized brass goggles pushed up on her forehead, works at a cluttered wooden workbench full of gears, springs and wires. She pulls the goggles down over her eyes, leans in and turns a tiny screwdriver in a clockwork gadget; it sparks and whirs, she flinches back, pushes the goggles up and laughs with delight. Warm lamp light, shallow depth of field, photographic realism, a clean face. Sound: ticking gears, a crackling spark and a delighted laugh. output: url: assets/videos/t2v_sample07_1152x640.mp4
text: >- A 3D animated scene in a warm, polished feature-animation style: a small round fox cub with bright orange fur, a cream belly, oversized green eyes and a big fluffy tail sits on a mossy log in a sunlit forest clearing. A blue butterfly drifts down and lands on the tip of its nose; the cub goes cross-eyed watching it, wrinkles its nose and sneezes, the butterfly flutters away, and the cub tumbles backwards off the log, then pops up again with leaves stuck in its fur and a big grin. Soft global illumination, saturated colours, smooth squash-and-stretch animation, gentle camera follow. Sound: birdsong, rustling leaves, a tiny sneeze and a happy squeak. output: url: assets/videos/t2v_sample11_1152x640.mp4
text: >- A young woman with freckles and windswept auburn hair stands by a tall window in soft daylight. A breeze blows strands of hair across her face; she tucks them behind her ear, turns her head from the window to look straight into the camera and smiles softly, then laughs, her eyes still on the camera to the last frame. Shallow depth of field, photographic realism, sharp detail. Sound: a gentle breeze and distant street noise. output: url: assets/videos/i2v_sample05_1152x640.mp4
text: >- Live-action documentary footage shot on 35 mm film: a real young inventor, not animated, not 3D, not CGI, with oversized brass goggles pushed up on her forehead, works at a cluttered wooden workbench full of gears, springs and wires. She pulls the goggles down over her eyes, leans in and turns a tiny screwdriver in a clockwork gadget; it sparks and whirs, she flinches back, pushes the goggles up and laughs with delight. Warm lamp light, shallow depth of field, photographic realism, a clean face. Sound: ticking gears, a crackling spark and a delighted laugh. output: url: assets/videos/i2v_sample07_1152x640.mp4
text: >- A 3D animated scene in a warm, polished feature-animation style: a small round fox cub with bright orange fur, a cream belly, oversized green eyes and a big fluffy tail sits on a mossy log in a sunlit forest clearing. A blue butterfly drifts down and lands on the tip of its nose; the cub goes cross-eyed watching it, wrinkles its nose and sneezes, the butterfly flutters away, and the cub tumbles backwards off the log, then pops up again with leaves stuck in its fur and a big grin. Soft global illumination, saturated colours, smooth squash-and-stretch animation, gentle camera follow. Sound: birdsong, rustling leaves, a tiny sneeze and a happy squeak. output: url: assets/videos/i2v_sample11_1152x640.mp4
text: >- A young woman with freckles and windswept auburn hair stands by a tall window in soft daylight. A breeze blows strands of hair across her face; she tucks them behind her ear, turns her head from the window to look straight into the camera and smiles softly, then laughs, her eyes still on the camera to the last frame. Shallow depth of field, photographic realism, sharp detail. Sound: a gentle breeze and distant street noise. output: url: assets/videos/t2v_sample05_640x1152.mp4
text: >- Live-action documentary footage shot on 35 mm film: a real young inventor, not animated, not 3D, not CGI, with oversized brass goggles pushed up on her forehead, works at a cluttered wooden workbench full of gears, springs and wires. She pulls the goggles down over her eyes, leans in and turns a tiny screwdriver in a clockwork gadget; it sparks and whirs, she flinches back, pushes the goggles up and laughs with delight. Warm lamp light, shallow depth of field, photographic realism, a clean face. Sound: ticking gears, a crackling spark and a delighted laugh. output: url: assets/videos/t2v_sample07_640x1152.mp4
text: >- A 3D animated scene in a warm, polished feature-animation style: a small round fox cub with bright orange fur, a cream belly, oversized green eyes and a big fluffy tail sits on a mossy log in a sunlit forest clearing. A blue butterfly drifts down and lands on the tip of its nose; the cub goes cross-eyed watching it, wrinkles its nose and sneezes, the butterfly flutters away, and the cub tumbles backwards off the log, then pops up again with leaves stuck in its fur and a big grin. Soft global illumination, saturated colours, smooth squash-and-stretch animation, gentle camera follow. Sound: birdsong, rustling leaves, a tiny sneeze and a happy squeak. output: url: assets/videos/t2v_sample11_640x1152.mp4
text: >- A young woman with freckles and windswept auburn hair stands by a tall window in soft daylight. A breeze blows strands of hair across her face; she tucks them behind her ear, turns her head from the window to look straight into the camera and smiles softly, then laughs, her eyes still on the camera to the last frame. Shallow depth of field, photographic realism, sharp detail. Sound: a gentle breeze and distant street noise. output: url: assets/videos/i2v_sample05_640x1152.mp4
text: >- Live-action documentary footage shot on 35 mm film: a real young inventor, not animated, not 3D, not CGI, with oversized brass goggles pushed up on her forehead, works at a cluttered wooden workbench full of gears, springs and wires. She pulls the goggles down over her eyes, leans in and turns a tiny screwdriver in a clockwork gadget; it sparks and whirs, she flinches back, pushes the goggles up and laughs with delight. Warm lamp light, shallow depth of field, photographic realism, a clean face. Sound: ticking gears, a crackling spark and a delighted laugh. output: url: assets/videos/i2v_sample07_640x1152.mp4
text: >- A 3D animated scene in a warm, polished feature-animation style: a small round fox cub with bright orange fur, a cream belly, oversized green eyes and a big fluffy tail sits on a mossy log in a sunlit forest clearing. A blue butterfly drifts down and lands on the tip of its nose; the cub goes cross-eyed watching it, wrinkles its nose and sneezes, the butterfly flutters away, and the cub tumbles backwards off the log, then pops up again with leaves stuck in its fur and a big grin. Soft global illumination, saturated colours, smooth squash-and-stretch animation, gentle camera follow. Sound: birdsong, rustling leaves, a tiny sneeze and a happy squeak. output: url: assets/videos/i2v_sample11_640x1152.mp4

LTX-2.5 Stage 1: 2-Step Distillation LoRA

Release candidate (checkpoint 12000) - training continues. Experimental! Good for quick previews, not production videos. My current plan is to reach 24K, as it is expensive to run this on H200, unless someone from HF or LTX wants to sponsor me :)

A LoRA for the distilled LTX-2.5 that takes stage 1 from 8 steps down to 2, picture and sound together, for text to video and image to video.

The included ComfyUI workflows combine speed improvements at every stage:

Stage 1 render: this LoRA, 2 steps instead of 8.
Refining: NVIDIA's SoL-Refiner-LTX-2.3-One-Step, 1 step instead of 3.
VAE decoding: PrunaAI's pruned, faster decoder, P-LTX-2.3-VAED.

This LoRA is for stage 1 only

LTX-2.5 renders in two stages: a half-size pass that composes the clip and its soundtrack, then a refine pass at full size. This LoRA covers the first pass. Do not apply it to the refiner.

What I use with it

Part	Model	Settings
Stage 1	LTX-2.5 distilled + this LoRA, strength 1.0	2 steps, sigmas `1.0, 0.909375, 0.0`, Euler Ancestral, CFG 1, half size
Refining (stage 2)	NVIDIA's one-step SoL-Refiner-LTX-2.3-One-Step	x2 latent upsample, AdaIN to stage 1, 1 step, sigmas `0.725, 0.0`, CFG 1
VAE decode	PrunaAI's pruned LTX-2.3 decoder, P-LTX-2.3-VAED	replaces the standard decoder

Examples

Stage 1 with this LoRA at 2 steps, then LTX-2.5's own 3-step refiner, on Comfy Cloud. 5 seconds with sound. Click a video to open it. The image to video clips start from the first frames in assets/i2v_images, made with Krea 2 Turbo.

ComfyUI workflows

The ComfyUI workflow

	Local ComfyUI	Comfy Cloud
Text to video	t2v local	t2v cloud
Image to video	i2v local	i2v cloud

Each workflow lists its models in a note. The refiner and the decoder are used through community ComfyUI conversions: the SoL-2.3 int8 transformer by szwagros and the Pruna decoder by Kijai. The refiner's prompt goes through ComfyUI's standard Gemma 3 12B and LTX-2.3 text projection. The AdaIN node comes from ComfyUI-LTXVideo. On Comfy Cloud, import this LoRA, the SoL-2.3 transformer and the Pruna decoder into your model library once, from their URLs.

Training

Trained with DMD2 on a single NVIDIA H200. This checkpoint is 12000 updates at rank 64. The teacher is LTX-2.5 distilled at its standard 8 steps for stage 1. The LoRA learns from distribution matching against the teacher, an adversarial critic on teacher clips and real footage, and regression on teacher clips. It covers the picture, the sound and the cross-attention between them, for text to video and image to video. It was trained on two resolutions, 576x320 and 320x576 at stage 1, which give 1152x640 and 640x1152 after refining, at 5 seconds. You can render at other resolutions, but this is what it is trained on.

Real footage and captions from OpenVid-1M were used in training, with thanks to its authors.

Files

File
`ltx25_stage1_distill_2step_rank_64_lora_comfyui.safetensors`	the LoRA, rank 64, ComfyUI keys
`ltx25_stage1_distill_2step_rank_64_lora.safetensors`	the LoRA, rank 64, diffusers keys

Licence

This LoRA is a derivative of LTX-2.5 and is distributed under the LTX-2.x Community License Agreement, a full copy of which is included, including the use restrictions of its Section 4 and Attachment A. The SoL-2.3 repository declares no licence of its own; it is a fine-tune of LTX-2.3, so check the LTX-2.x terms before using it commercially.