COMMUNITY PAGE
Run LTX 2.5 now on Floyo
Home / Model / LTX 2.5 on Floyo
AI VIDEO GENERATION
Run LTX 2.5 on Floyo
Lightricks' open-weight video model with Diffusion Fidelity Rendering. Faster-than-real-time generation (6.8 seconds for a 10-second clip). Native multi-shot, native 4K HDR, RAW EXR workflow, auto-duration, Gemma 4 12B text encoder, and the fewest visual artifacts of any model tested. Open weights, fine-tunable, 16GB VRAM minimum.
Run Lightricks' LTX 2.5 through ComfyUI in your browser. No API key, no installs, no local GPU.
|
Speed (On-Prem) 6.8s for 10s clip (faster than real time) |
Resolution 720p / 1080p / 1440p / Native 4K HDR |
|
Artifacts 0.28 per clip (#1 of 10 models tested) |
Min VRAM 16GB (open weights) |
No installation. Runs in browser. Updated August 2026.
What You Get
LTX 2.5 is Lightricks' latest open-weight video model with Diffusion Fidelity Rendering, a new technology that allocates rendering compute by scene complexity for cinema-grade pixel quality that holds up frame by frame. Generates a 10-second clip in 6.8 seconds on-prem (faster than real time). Native multi-shot scenes with consistent characters, lighting, and voice across cuts. Gemma 4 12B text encoder for stronger prompt adherence. Auto-duration matching. Native 4K HDR output. RAW EXR export for professional color grading. Fewest visual artifacts of any model in 10-model testing (0.28 per clip). Two API tiers: Fast ($0.09/s at 720p) and Pro ($0.12/s at 720p). Open weights on HuggingFace. Fine-tunable. Runs on 16GB VRAM. Available as ComfyUI nodes on Floyo with 2 workflows.
LTX 2.5 WORKFLOWS ON FLOYO
i2v
LTX 2.5
Video
Animate any image into a video clip with synced audio using LTX 2.5, Lightricks' 22B open-weights video model. Upload a frame, describe the motion, hit run.
LTX 2.5 for Image to Video
Animate any image into a video clip with synced audio using LTX 2.5, Lightricks' 22B open-weights video model. Upload a frame, describe the motion, hit run.
ltx 2.5
open source
text to video
Video
video generation
Generate video with synchronized audio from a text prompt using LTX-2.5, the 22B open-weights model by Lightricks. Write a scene, hit run, get a clip with sound.
LTX 2.5 for Text to Video
Generate video with synchronized audio from a text prompt using LTX-2.5, the 22B open-weights model by Lightricks. Write a scene, hit run, get a clip with sound.
What is LTX 2.5?
LTX 2.5 is Lightricks' newest open-weight video generation model and the latest in the LTX series (following LTXV, LTX-2, and LTX-2.3). It introduces Diffusion Fidelity Rendering (DFR), a new video generation technology that allocates rendering compute proportionally to scene complexity. Simple scenes render fast. Complex scenes with fine textures, intricate motion, and dense detail receive more compute. The result is cinema-grade pixel quality that holds up frame by frame on a large screen.
Speed is the headline number. Self-hosted on 2x GB200 GPUs, LTX 2.5 generates a 10-second 720p video clip in 6.8 seconds. That is faster than real time. Via the API on fal.run (including queue time), it generates in 23.7 seconds at 1080p. For comparison: Gemini Omni Flash takes 52 seconds, MiniMax H3 takes 180 seconds, Seedance 2.0 takes 196 seconds, and Kling 3.0 Pro takes 398 seconds for the same duration.
Visual artifact count is the other standout metric. In a 10-model comparison across 98 text-to-video prompts, LTX 2.5 Pro scored 0.28 glitches per clip (blotches, broken texture, melted/smeared areas). FLUX 3 scored 0.45. MiniMax scored 0.46. Seedance 2.5 scored 0.69. Veo 3.1 scored 1.20. LTX 2.5 produces the cleanest frames of any model tested.
The model uses Gemma 4 12B as its text encoder (replacing T5 from earlier versions) paired with a custom prompt enhancer. This is why it follows complex creative instructions from shorter, simpler prompts. Auto-duration prediction generates the right clip length for the requested action, so a 3-second gesture is not padded to fill 10 seconds.
On Floyo, LTX 2.5 runs through ComfyUI nodes on H100 NVL GPUs. Two workflows cover text-to-video and image-to-video. No weight downloads, no VRAM management, no local setup.
What are LTX 2.5's technical specifications?
LTX 2.5 uses Diffusion Fidelity Rendering with a new diffusion video decoder. Gemma 4 12B text encoder with custom prompt enhancer. Native multi-shot with character, environment, lighting, and voice consistency. Auto-duration. Native 4K HDR and RAW EXR export. 6.8-second generation for a 10-second clip on-prem. 0.28 artifacts per clip (#1 of 10 models). Open weights. 16GB VRAM minimum. Two API tiers: Fast and Pro.
| Spec | Details |
|---|---|
| Developer | Lightricks (LTX) |
| Key Innovation | Diffusion Fidelity Rendering (DFR): compute allocation scaled to scene complexity |
| Text Encoder | Gemma 4 12B + custom prompt enhancer |
| Video Decoder | New diffusion video decoder (cleaner motion, fewer artifacts) |
| Resolution | 720p, 1080p, 1440p, 4K (native HDR) |
| Frame Rate | 24fps |
| Multi-Shot | Native (character, environment, lighting, voice consistent across cuts) |
| Auto Duration | Generates right clip length for requested action |
| RAW Export | EXR output for professional color grading and finishing |
| Video Editing | Beta (edit real footage with instructions) |
| Artifact Score | 0.28 per clip (Pro), 0.39 (Fast) — #1 of 10 models tested |
| Speed (On-Prem) | 6.8 seconds for 10s clip at 720p (2x GB200) |
| Speed (API) | 23.7 seconds at 1080p (fal.run, includes queue) |
| API Tiers | Fast: $0.09/s (720p), $0.13/s (1080p), $0.30/s (4K). Pro: $0.12/s (720p), $0.17/s (1080p). |
| Min VRAM | 16GB (open weights) |
| Open Weights | Yes (HuggingFace: Lightricks/LTX-2.5). Fine-tunable. Free under $10M ARR. |
| Deployment | On-prem, edge, API (runs on any GPU) |
| ComfyUI Access | Native support on Floyo (2 workflows) |
| Family | LTXV → LTX-2 → LTX-2.3 → LTX 2.5 |
What can you create with LTX 2.5?
LTX 2.5 covers cinematic video generation, image animation, multi-shot scene creation, 4K HDR production, RAW color grading workflows, video editing (beta), VFX production, marketing content, gaming cutscenes, and physical AI simulation. The faster-than-real-time speed makes it suited for interactive workflows, real-time applications, and high-volume production where every other model is too slow.
| Capability | What It Does | Use Case |
|---|---|---|
| Faster-Than-Real-Time | Generate a 10-second video in 6.8 seconds on-prem. 23.7 seconds via API at 1080p. The fastest video model benchmarked across 10 competitors. | Real-time applications, bulk production, interactive workflows |
| Native Multi-Shot | Generate connected shots that hold character, environment, lighting, and voice consistent across cuts. A multi-shot scene is one generation, not stitched clips. | Short films, ad sequences, brand stories, narrative content |
| Native 4K HDR | Generate high-resolution HDR footage for professional finishing. No upscaling pass needed. Direct from model to delivery at cinema resolution. | Broadcast, cinema, large-format display, VFX plates |
| RAW EXR Workflow | Generate and edit inside professional color and finishing pipelines. Export as EXR without compromising the master. Full creative control in DaVinci Resolve, Nuke, or ACES. | Film post-production, color grading, VFX compositing |
| Fine-Tuning | Open pretrained base checkpoint built for fine-tuning. Adapt to your domain, data, and IP. Deploy on your own infrastructure. Free under $10M ARR. | Custom models, brand-specific video, domain specialization |
| Pipeline Integration | Chain with image models in ComfyUI. Generate a character with Nano Banana or Ideogram V4, animate with LTX 2.5, add voiceover with ElevenLabs, export as 4K HDR EXR. | Multi-model production pipelines |
What are LTX 2.5's key features?
LTX 2.5 is built around three design goals: speed, visual fidelity, and openness. Diffusion Fidelity Rendering delivers cinema-grade pixels. Faster-than-real-time inference unlocks interactive and real-time applications. Open weights with fine-tuning and 16GB VRAM minimum make it deployable on hardware you control. Everything else follows from these three.
Diffusion Fidelity Rendering (DFR)
A new video generation technology that allocates rendering compute proportionally to scene complexity. A simple scene with a static background renders fast. A complex scene with fine fabric textures, flowing water, and intricate hair movement receives more compute automatically. The result is pixel quality that holds up on a cinema screen, frame by frame. This is why LTX 2.5 Pro scored 0.28 artifacts per clip, the lowest of any model tested.
Fastest Generation in 10-Model Testing
On-prem at 720p: 6.8 seconds for a 10-second clip (2x GB200). Via API at 1080p: 23.7 seconds. For context: Gemini Omni Flash takes 52 seconds. Grok 1.5 takes 63 seconds. Veo 3.1 takes 70 seconds. MiniMax H3 takes 180 seconds. Seedance 2.5 takes 317 seconds. Kling 3.0 Pro takes 398 seconds. LTX 2.5 renders at a higher resolution than most models in this comparison and is still the fastest.
Native Multi-Shot Consistency
Generate connected shots that hold character identity, environment design, lighting setup, and voice consistent across scene cuts. This is one generation pass, not post-processed clip stitching. A dialogue scene with two characters in three camera angles renders as a coherent sequence with matching lighting, wardrobe, and spatial relationships in every shot.
Gemma 4 12B Text Encoder
Replaces T5 from earlier LTX versions. Paired with a custom prompt enhancer. This is why LTX 2.5 follows complex creative instructions from shorter, simpler prompts. You do not need to write a paragraph to get a specific shot. A concise description with cinematographic intent is enough. The model interprets creative direction, not just keyword matching.
Native 4K HDR + RAW EXR Export
Generate at 4K with HDR natively. No upscaling pass. For professional finishing, export as EXR and work inside DaVinci Resolve, Nuke, or any ACES-compatible color pipeline without compromising the master. This is the first open-weight video model to support a professional RAW workflow from generation to delivery.
Open Weights, Fine-Tunable, 16GB VRAM
Open weights on HuggingFace (Lightricks/LTX-2.5). Runs on any GPU with 16GB VRAM. No mandatory branding. Fine-tune on your own data and IP. Deploy on-prem, at the edge, or through the API. Free for commercial use under $10M ARR. This is the most permissive license and lowest hardware requirement among production-grade video models.
How does LTX 2.5 compare to other video models?
LTX 2.5 leads on speed (6.8s on-prem, fastest of 10 models), visual cleanliness (0.28 artifacts/clip, #1), open-weight accessibility (16GB VRAM, fine-tunable, no mandatory branding), and professional output formats (native 4K HDR, RAW EXR). MiniMax H3 leads on reference input capacity. Wan 3.0 leads on duration (30 seconds). HappyHorse 1.0 leads on arena Elo. LTX 2.5's edge: the fastest, cleanest, and most deployable open video model available.
| Model | Speed (10s I2V) | Artifacts/Clip | Open Weights | Max Resolution |
|---|---|---|---|---|
| LTX 2.5 | 6.8s (on-prem) | 0.28 (Pro, #1) | Yes (16GB, fine-tune) | Native 4K HDR |
| MiniMax H3 | 180s | 0.46 | No | Native 2K |
| Seedance 2.5 | 317s | 0.69 | No | 720p |
| Veo 3.1 | 70s | 1.20 | No | 720p |
| Kling 3.0 Pro | 398s | 0.76 | No | ~1080p |
Source: LTX official benchmarks (ltx.io/model/ltx-2-5), speed measured via fal.run (I2V, 10s, 24fps), artifact scoring across 98 T2V prompts with automated grading. HuggingFace model card (Lightricks/LTX-2.5). August 2026.
How does LTX 2.5 work?
LTX 2.5 uses Diffusion Fidelity Rendering with a new diffusion video decoder that allocates rendering compute by scene complexity. The Gemma 4 12B text encoder processes your prompt with full contextual understanding, paired with a custom prompt enhancer that enriches creative intent. The model generates video frames with a distilled inference pipeline trained using reinforcement learning for quality at reduced compute.
The new diffusion video decoder is what produces cleaner motion and fewer artifacts compared to LTX 2.3. It renders each frame with awareness of its complexity: areas with fine texture, rapid motion, or intricate detail receive proportionally more compute. Static backgrounds render fast. Complex fabric, hair, water, and particle effects receive the rendering time they need.
Auto-duration prediction estimates the right clip length based on the described action. A prompt describing a quick gesture generates a short clip. A prompt describing a complex sequence generates a longer one. This eliminates the "padded" feeling where short actions get stretched to fill a fixed duration. You can override and set exact duration manually.
On Floyo, LTX 2.5 runs through ComfyUI nodes on H100 NVL GPUs. Your prompt (and optional first-frame image for I2V) feeds through the model pipeline, and the generated video returns to your ComfyUI canvas. Chain with other models: generate a character with Nano Banana, animate with LTX 2.5, add voiceover with ElevenLabs, export as 4K HDR.
Fair warning: LTX 2.5 is open-weight with a permissive license (free under $10M ARR, no mandatory branding). Video editing is in beta. The speed benchmarks use 2x GB200 GPUs for on-prem numbers; consumer hardware will be slower. The artifact scoring is automated, not human-rated, and covers 98 prompts (preliminary results, expected to evolve). RAW EXR export is a professional feature that requires compatible post-production software. API pricing applies through your Floyo API Wallet.
Frequently Asked Questions
Common questions about running LTX 2.5 on Floyo.
You can start with Floyo's free pricing plan. To continue using the service beyond the free tier, upgrade your Floyo pricing plan. LTX 2.5 is open-weight, so there is no additional API cost beyond your Floyo plan when running through native ComfyUI nodes.
Open Floyo in your browser, search "LTX 2.5" in the template library, and pick the text-to-video or image-to-video workflow. Click Run, write your prompt, and generate. Floyo handles the H100 GPU, ComfyUI environment, and model weights. No local install, no 16GB VRAM requirement, no Python setup.
Lightricks, an AI company behind the LTX model family and LTX Studio. LTX 2.5 is the latest in the series (LTXV, LTX-2, LTX-2.3, LTX 2.5). Open weights on HuggingFace (Lightricks/LTX-2.5). Code on GitHub (Lightricks/LTX-2). Permissive license, free under $10M ARR.
A new video generation technology in LTX 2.5 that allocates rendering compute proportionally to scene complexity. Simple scenes render fast. Complex scenes with fine textures, rapid motion, and dense detail receive more compute. The result is consistent pixel quality that holds up frame by frame, even on a large cinema screen.
The fastest benchmarked. On-prem (2x GB200): 6.8 seconds for a 10-second clip. Via API: 23.7 seconds at 1080p. For comparison: Gemini Omni Flash 52s, Grok 1.5 63s, Veo 3.1 70s, MiniMax H3 180s, Seedance 2.0 196s, Seedance 2.5 317s, Kling 3.0 Pro 398s. LTX renders at 1080p (higher than most) and is still the fastest.
Yes. LTX 2.5 includes a pretrained base checkpoint built for fine-tuning. Adapt it to your domain, data, and IP. Deploy on your own infrastructure. No mandatory branding. Free under $10M annual revenue. The community has already produced LoRAs for styles, camera motion, outpainting, greenscreen avatars, and more.
Yes. Floyo runs ComfyUI, which lets you chain multiple models. Generate a character with Nano Banana or Ideogram V4, animate with LTX 2.5, add voiceover with ElevenLabs or Fish Audio S2, upscale with Topaz. Or export as 4K HDR EXR for professional color grading in DaVinci Resolve.
Yes. LTX 2.5 uses open weights with a permissive license. No mandatory branding. Free commercial use for companies under $10M ARR. Enterprises above that threshold should contact Lightricks for licensing. Generated videos can be used for products, marketing, client work, and any commercial context.
Try LTX 2.5 on Floyo
Lightricks' open-weight video model with Diffusion Fidelity Rendering. Faster than real time. Fewest artifacts. Native 4K HDR. RAW EXR export. Fine-tunable on 16GB VRAM. Run it in your browser.
Try LTX 2.5 Now → Browse All ModelsRelated Reading
Film and Animation Workflows on Floyo
VFX and Post Production on Floyo
Last updated: August 2026. Specs from LTX official model page (ltx.io/model/ltx-2-5), HuggingFace model card (Lightricks/LTX-2.5), GitHub (Lightricks/LTX-2), LTX API pricing documentation, speed benchmarks measured on fal.run, artifact scoring across 98 T2V prompts with automated grading, and LTX community library.
Run LTX 2.5 online through ComfyUI on Floyo. Lightricks' open-weight video model with Diffusion Fidelity Rendering. Faster than real time (6.8s for 10s clip). Fewest artifacts of any model tested (0.28/clip). Native multi-shot, 4K HDR, RAW EXR export, Gemma 4 12B encoder, auto-duration. Open weights, fine-tunable, 16GB VRAM. 2 workflows. No install, no GPU, browser-based. Free to try.