Multi‑scene — stays coherent across scene cuts / several shots within one clip.
Audio — talking scenes generate synced audio.
Talking characters — much better faces, mouth / lip‑sync and expression for people speaking to camera.
Higher quality — sharper, more detailed, more photorealistic output.

Best paired with the render‑to‑real endpoint's enable_detail_refine option (extra 2× spatial‑upscale + detailer pass). File: 3DREAL-strong-v2.safetensors.

Made with Strong v2 — a 1930s chase short where every shot starts as a Blender blockout (action, cameras and cuts owned by the 3D scene):

IRON CANYON — generated with 3DREAL Strong v2 from a blocky Blender scene

Full write-up of the pipeline behind this film: From clay 3D render to a real action short

This is an IC‑LoRA (in‑context LoRA) for LTX‑2.3 that turns a rough grey 3D viewport animation (a simple low‑poly blockout) into a fully photorealistic, film‑quality video, while keeping the exact composition, camera move and layout of the input.

It is made for 3D scenes and synthetic data — Blender blockouts, game‑engine viewports and other CG / synthetic renders become photoreal video.

Examples

Top = generated result · Bottom = 3D render input

How to use

The easiest way is the ready‑to‑use endpoint on fal (the adapter is already loaded, nothing to copy):

https://fal.ai/models/fal-ai/ltx-2.3-quality/render-to-real

python

import fal_client

result = fal_client.subscribe(
    "fal-ai/ltx-2.3-quality/render-to-real",
    arguments={
        "video_url": "<your 3D / CG / game render>",
        "image_url": "<optional photoreal reference for the first frame>",
        # the prompt always keeps the 3DREAL trigger; describe the photoreal result
        "prompt": "3DREAL. Make it photorealistic. A cargo ship stacked with shipping containers in a busy harbor.",
        "resolution": "720p",
    },
)
print(result["video"]["url"])

Trigger word: 3DREAL — always kept at the start of the prompt
Inputs: a 3D / CG / game render (video_url) + an optional photoreal reference for the first frame (image_url)
Base model: LTX‑2.3

Variants — Light, Strong & Strong v2

This repo ships three adapters — pick based on how far you want to push the look:

3DREAL Light (3DREAL-light.safetensors) — faithful, gentle transformation. Stays close to the input's structure, composition and motion; fewer hallucinations. This is the default used by the hosted render-to-real endpoint.
3DREAL Strong (3DREAL-strong.safetensors) — stronger, more aggressive photoreal push. More realism and detail, often better on busy/complex scenes, but may drift a little more from the input.
3DREAL Strong v2 (3DREAL-strong-v2.safetensors) — new, improved. Multi‑scene coherence (holds across scene cuts / multiple shots), audio generation, much better on talking characters (faces, mouth / lip motion, expression), and higher overall quality & detail. Pairs best with the endpoint's enable_detail_refine option for sharper, higher‑resolution output. See the example GIF in the update section above.

To run a specific variant with explicit weights, use the LoRA endpoint:

python

import fal_client

result = fal_client.subscribe(
    "fal-ai/ltx-2.3-quality/reference-video-to-video/lora",
    arguments={
        "video_url": "<your 3D / CG / game render>",
        "image_url": "<photoreal reference for the first frame>",
        "prompt": "3DREAL. Make it photorealistic. <your scene>.",
        "loras": [{"path": "<Light or Strong .safetensors URL>", "scale": 1.0}],
        "resolution": {"width": 1280, "height": 704},
    },
)
print(result["video"]["url"])

Weights:

Light — https://v3b.fal.media/files/b/0a9fd993/oow7bW3VueUPTukqAQkd7_3DREAL.safetensors
Strong — https://v3b.fal.media/files/b/0a9fe083/H7caCyG_wt9hy_51tMEmu_3DREAL-strong.safetensors
Strong v2 (new) — https://huggingface.co/fal/LTX-2.3-3DREAL-LoRA/resolve/main/3DREAL-strong-v2.safetensors

Example prompts

text

3DREAL. Make it photorealistic. A cargo ship stacked with shipping containers in a busy harbor at dawn.
3DREAL. Make it photorealistic. A steam train winding through a misty mountain valley.
3DREAL. Make it photorealistic. A lone car driving down an endless desert highway at sunset.
3DREAL. Make it photorealistic. A grand ballroom with crystal chandeliers, slow camera glide.
3DREAL. Make it photorealistic. A coastal cliff with a lighthouse and crashing waves at golden hour.
3DREAL. Make it photorealistic. A sleek bullet train speeding across green countryside.

Created by Lovis Odin (@OdinLovis) with fal (@fal · fal.ai)

fal/LTX-2.3-3DREAL-LoRA

LTX‑2.3 IC‑LoRA — 3D render to Photoreal

▶ Run it directly on fal (no setup, LoRA built in)

🆕 Update — v2 (3DREAL-strong-v2.safetensors)

Examples

How to use

Variants — Light, Strong & Strong v2

Example prompts

fal/
LTX-2.3-3DREAL-LoRA

🆕 Update — v2 (`3DREAL-strong-v2.safetensors`)