Overview

LTX-2.3-3DREAL-LoRA is an in-context LoRA (incontext Low-Rank Adaptation) for LTX-2.3 that transforms 3D renders into photorealistic, film-quality videos while preserving the exact composition, camera movement, and layout of the input. Created by fal, this model specializes in converting synthetic content—Blender blockouts, game engine viewports, and other CG renders—into photoreal video output. The model operates with a trigger word 3DREAL that anchors the prompt and ships in two variants: a Light version (faithful, gentle transformation with fewer hallucinations) and a Strong version (aggressive photoreal push with more detail but potential drift from input structure). The easiest deployment path uses fal's hosted endpoint at https://fal.ai/models/fal-ai/ltx-2.3-quality/render-to-real, which has the LoRA pre-loaded and requires no local setup. The model accepts video input (video_url), an optional photoreal reference image for the first frame (image_url), and text prompts that automatically prepend the 3DREAL trigger word. Output resolution defaults to 720p via the quick endpoint, though the underlying LoRA endpoint supports explicit resolution specifications up to 1280x704 pixels.

Best use cases

Converting Blender blockouts to photoreal shots. This model handles the specific challenge of turning low-poly geometric layouts into believable photorealistic footage while maintaining the blocking and camera work an animator or designer established. The Light variant preserves composition faithfully, making it ideal for production pipelines where the 3D layout is final and only the rendering needs replacement. This is particularly valuable in previsualization workflows where directors need to see blocking decisions transformed into believable lighting and materials without re-rendering in a heavyweight rendering engine.

Game engine viewport to cinematic video. Unreal Engine, Unity, or other game engine outputs often contain rough geometry and placeholder materials. This model converts those technical renders into cinematic quality by understanding the spatial composition from the 3D input and synthesizing photorealistic surfaces, lighting, and atmospheric effects. The preserved camera movement means any dynamic shots or fly-throughs maintain their original motion while gaining photorealism, useful for creating marketing material or cinematics from game assets.

Synthetic data augmentation with photorealism. When you have 3D-generated data with perfect labels (depth, geometry, object masks) but need photorealistic appearance for training computer vision models, this LoRA bridges that gap. The Strong variant adds realistic detail and materials while the underlying 3D structure remains unchanged, creating hybrid training data that combines synthetic precision with natural appearance. This applies to autonomous vehicle training, robotics simulation, or any domain where labeled synthetic data needs visual believability.

Architectural visualization with motion. Static renders of buildings and interiors can be transformed into walkthrough or fly-through videos with photorealistic materials, lighting, and reflections. The model preserves the camera path while synthesizing realistic textures, glass reflections, and environmental lighting that might take hours to render traditionally. Real estate, architectural presentation, and interior design firms benefit from the speed compared to traditional rendering engines.

VFX integration and plate replacement. When you have a simple 3D geometry layer or blockout that needs to match photorealistic plates, this model can generate photorealistic content that respects the underlying 3D structure and camera movement. This is useful for visual effects work where you need fast iterations on photoreal objects within a live-action scene context.

Limitations

Structural faithfulness trades against realism in the Strong variant. The Light variant stays close to input composition with fewer hallucinations, but the Strong variant may drift from the input geometry and layout when pushing for photorealism. Complex or ambiguous 3D blockouts may receive synthetic details that don't match the original intent, requiring careful prompt engineering and optional reference images to guide the output.

Requires high-quality 3D input. The model depends on receiving a clean, coherent 3D render or blockout. Broken geometry, extremely low-poly scenes with ambiguous topology, or severely compressed video artifacts degrade results. The input should represent a valid spatial layout that the model can reasonably interpret as a scene.

Limited to 720p on the default endpoint. The ready-to-use hosted endpoint produces 720p output. The underlying LoRA endpoint supports higher resolution (1280x704), but this requires manual LoRA weight loading and API configuration, adding deployment complexity.

Motion preservation depends on clean input. While the model maintains camera movement from the input, extremely fast cuts, motion blur-heavy footage, or video artifacts can confuse the motion tracking. The reference image feature (image_url) helps ground the first frame but doesn't control subsequent frames directly.

Inference speed not specified. The README provides no details on generation time, GPU memory requirements, or throughput. Users deploying locally must test latency requirements against their infrastructure without baseline expectations.

Limited customization without reference images. The 3DREAL trigger word is always preserved and cannot be modified. Fine-tuning control relies on prompt engineering and optional first-frame reference images rather than granular parameters. The LoRA scale can be adjusted (0.0 to 1.0+) but the model provides no guidance on optimal settings for different scene types.

License classified as "other". The repository does not specify a standard open-source or commercial license. Users must contact fal directly to clarify usage rights and commercial deployment terms.

How it compares

vs. ltx-2.3-quality/render-to-real — This model IS the 3DREAL LoRA variant of that endpoint. The standalone render-to-real endpoint is the hosted version with the Light LoRA already loaded and optimized for ease of use without setup. Choose the direct endpoint if you want a one-line API call; choose the LoRA repository if you need the Strong variant or want to load weights locally and customize the scale.

vs. ltx-2.3-quality/image-to-video/lora — The image-to-video LoRA generates video from static images with custom LoRA adaptation for general stylization. The 3DREAL LoRA is purpose-built for 3D-to-photo transformation with the specific goal of preserving spatial structure while synthesizing photorealism. Choose 3DREAL if your input is 3D/CG renders; choose image-to-video if you need to animate or extend static photographs or artwork.

vs. ltx-2.3-22b/image-to-video/lora — The 22B model is larger and capable of more general video generation from images with custom LoRA. The 3DREAL model is smaller and specialized for 3D-render-to-photo conversion with two variants (Light/Strong) tuned for that specific domain. Use 22B if you need maximum quality for general video generation; use 3DREAL if your input is 3D renders and you want composition preservation.

vs. ltx-2.3-quality/hdr/lora — The HDR LoRA enhances video with high dynamic range processing from a reference. The 3DREAL LoRA transforms 3D renders into photorealistic video. These serve different purposes: use HDR to enhance existing video; use 3DREAL to convert synthetic 3D content to photoreal video.

vs. ltx-2.3-22b/distilled/image-to-video/lora — The distilled model is a smaller, faster variant of the 22B for general image-to-video. The 3DREAL LoRA is specialized and optimized specifically for 3D-render conversion. Use distilled for fast, general-purpose video generation from images; use 3DREAL when you need domain-specific handling of synthetic 3D content with strict composition preservation.

Technical specifications

LTX-2.3-3DREAL-LoRA is built as a LoRA adapter for the LTX-2.3 base model. The repository ships two weight files:

• 3DREAL Light (3DREAL-light.safetensors) — Faithful, conservative transformation that stays close to input structure, composition, and motion with minimal hallucinations

• 3DREAL Strong (3DREAL-strong.safetensors) — Aggressive photoreal push offering more realism and detail, often better on complex scenes but with potential drift from input layout

The model operates on video input formatted as MP4 or similar standard video files. The hosted endpoint on fal supports 720p output by default. The underlying LoRA endpoint accepts explicit resolution specifications: width and height can be set to values such as 1280x704 pixels. Input video can be any standard frame rate and duration, though no limits are specified. The reference image (image_url) accepts standard image formats (JPEG, PNG) and anchors the first frame of output to a photorealistic reference when provided.

The LoRA weights are stored in .safetensors format, compatible with fal's LTX-2.3 endpoints via the LoRA loading mechanism. The model uses a scale parameter (typical range 0.0 to 1.0, potentially higher) to modulate adapter influence, allowing users to blend between the base model and the photoreal aesthetic. The 3DREAL trigger word is mandatory in all prompts and is automatically prepended by the hosted endpoint.

Model inputs and outputs

Inputs

• video_url (string, required): URL to a 3D render, CG video, game engine viewport, or Blender blockout output in standard video format

• image_url (string, optional): URL to a photorealistic reference image for the first frame, used to anchor appearance and guide photoreal synthesis

• prompt (string, required): Text description of the desired photorealistic result; the 3DREAL trigger word is automatically prepended and cannot be removed

• resolution (string or object, optional): Output resolution; hosted endpoint supports "720p" as default; LoRA endpoint accepts explicit {"width": 1280, "height": 704} format

• loras (array, optional): LoRA weights specification with path URL and scale factor (0.0 to 1.0+); required when using the LoRA endpoint to select Light or Strong variant

Outputs

• video (object): Generated video file with URL to MP4 or similar format; resolution matches input specification

• video.url (string): Direct URL to the generated photorealistic video output

Getting started

import fal_client

# Using the pre-loaded hosted endpoint (Light variant, simplest approach)
result = fal_client.subscribe(
    "fal-ai/ltx-2.3-quality/render-to-real",
    arguments={
        "video_url": "https://example.com/my-3d-render.mp4",
        "image_url": "https://example.com/first-frame-reference.jpg",
        "prompt": "3DREAL. Make it photorealistic. A cargo ship stacked with shipping containers in a busy harbor at dawn.",
        "resolution": "720p",
    },
)
print(result["video"]["url"])

For the Strong variant or custom weight loading:

import fal_client

# Using the LoRA endpoint with explicit Strong variant
result = fal_client.subscribe(
    "fal-ai/ltx-2.3-quality/reference-video-to-video/lora",
    arguments={
        "video_url": "https://example.com/my-3d-render.mp4",
        "image_url": "https://example.com/first-frame-reference.jpg",
        "prompt": "3DREAL. Make it photorealistic. A grand ballroom with crystal chandeliers.",
        "loras": [{"path": "https://v3b.fal.media/files/b/0a9fe083/H7caCyG_wt9hy_51tMEmu_3DREAL-strong.safetensors", "scale": 1.0}],
        "resolution": {"width": 1280, "height": 704},
    },
)
print(result["video"]["url"])

Frequently asked questions

Q: Should I use the Light or Strong variant?

A: Use Light if you need faithful preservation of the 3D input's composition, motion, and layout with minimal hallucinations—ideal for production pipelines where the blockout is final. Use Strong if your scene is complex or busy and you want maximum photorealistic detail and realism, accepting potential minor drift from the original geometry.

Q: Can I use this model for commercial projects?

A: The license is listed as "other" without standard terms. You must contact fal directly to clarify commercial usage rights and licensing terms before deploying to production.

Q: What GPU or hardware do I need to run this locally?

A: The README does not specify VRAM requirements, inference speed, or minimum hardware specifications. Test the hosted fal endpoint first to understand latency requirements, then contact fal for local deployment guidance.

Q: What resolution can I generate?

A: The hosted render-to-real endpoint produces 720p output. The underlying LoRA endpoint supports higher resolution specifications (e.g., 1280x704 pixels), but this requires manual LoRA loading and API configuration.

Q: How important is the reference image (image_url) parameter?

A: The reference image anchors the first frame to a photorealistic appearance and helps guide the model's style for the entire video. It is optional but recommended to ensure the output aesthetic matches your intent, especially when using the Strong variant.

Q: Can I modify or fine-tune the 3DREAL trigger word?

A: No. The 3DREAL trigger word is mandatory and automatically prepended to all prompts; it cannot be changed or removed. Control the output through prompt description and the reference image.

Q: How does this model preserve the camera movement and composition?

A: The model analyzes the input video's spatial layout, geometry, and motion vectors from the 3D render, then synthesizes photorealistic surfaces, lighting, and materials while respecting the original camera path and composition. The Light variant is tuned to stay closer to the input structure with fewer hallucinations.

Q: What types of 3D renders work best with this model?

A: Clean, coherent blockouts from Blender, game engines, or 3D software with clear geometric structure work best. Extremely low-poly, ambiguous geometry, or heavily compressed video artifacts degrade results. The input should represent a valid spatial scene that the model can reasonably interpret.

This is a simplified guide to an AI model called LTX-2.3-3DREAL-LoRA maintained by fal. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.