DiffusionGemma LTX-2.5 Target Profile

If you've ever watched an LLM produce a beautiful LTX prompt for a task LTX can't actually do - or jam a 90-second action sequence into a 5-second clip - you know the problem this node solves. DiffusionGemma can't guess what your target model wants. DiffusionGemma LTX-2.5 Target Profile is the conversation where you tell it: we're going to LTX, here's the mode, here's how long, here's how much audio to write, here's the camera skill ceiling. Everything else the Director does is downstream of that agreement.

It's one of the "choose exactly one" family of Target Profile nodes (the others target MiniMax H3 and Ideogram 4). You drop one between the Context Hub and the CoT Generator, and it emits a target_profile_config bundle plus a JSON mirror that constrains everything the prompt model writes. Pick the LTX profile for LTX, the H3 profile for H3 - they produce genuinely different prompt contracts.

The inputs that matter

• generation_mode - Auto (recommended) infers from your Context Hub wiring: no frame socket connected means text-to-video, an image connected means image-to-video (treated as the exact first frame), and first+last-frame requires both dedicated frame sockets. Override only for unusual graphs.
• target_duration_seconds - 0 auto-resolves: source media duration, then an explicit duration in your brief, then a 5-second planning default. This sets LTX shot, action, sound, and speech budgets - it's a planning budget, not a hard sampler clamp.
• audio_mode - auto_scene_audio (let LTX do its synchronized audio thing), explicit_sound_design (give it real direction), or visual_only. LTX-2's whole selling point is native audio+video in one pass, so this is where you decide how much of the prompt's audio budget Director spends on sound versus picture.
• audio_guidance - optional ambience, effects, music, speaker, delivery, or exact quoted wording to weave in.
• style_guidance - the visual treatment; keep it concrete and compatible with any conditioning frames.
• long_horizon_mode - experimental. Off (default) preserves the classic compiler. Auto (>20 seconds) activates only when the resolved duration is strictly above 20s. On forces it at any length. When active, Director silently plans four duration-scaled phases (establish, commit, sustain/reveal, settle/hold) and compiles them into one continuous paragraph of ≤200 words, four to eight sentences. These are not shots or chapters - no headings, no timecodes. It improves semantic pacing for long continuous takes but adds no geometric conditioning; pair it with first/last frames or a duration-matched depth/control sequence if you need the final composition constrained.
• camera_capability - Stable / base model (the safe default: locked framing or one restrained push/pull/pan/tilt/track) versus Advanced / controlled camera (preserves ambitious orbits, rolls, sweeping parallax - only with suitable motion control or a camera LoRA feeding LTX).

What comes out and where it goes

Outputs are target_profile_config and target_profile_json. Wire the config into the CoT Generator's target_profile_config input. Director then compiles the prompt, and the JSON Splitter turns it into the ltx_prompt / negative_prompt / resolution outputs that feed the gate and, ultimately, the native LTX conditioning nodes.

Two LTX-specific behaviors worth knowing. First, Director requires I2V and first+last-frame prompts to open every shot with one of LTX's six canonical shot types before describing the camera path, and it rejects cuts or overload that can't fit the duration - that's why a source video or conditioning frame matters: it supplies opening scale and viewpoint instead of making Director guess and block. Second, negative_prompt_mode defaults to auto, which lets Director author a genuinely useful LTX negative; empty forces none, custom uses your negative_prompt_guidance.

Installing it

One of 55 nodes in exportAnything/ComfyUI-DiffusionGemmaPromptBuilder:

cd ComfyUI/custom_nodes
git clone https://github.com/exportAnything/ComfyUI-DiffusionGemmaPromptBuilder

or via ComfyUI Manager (DiffusionGemma Prompt Builder), then restart. This node itself is pure configuration - the model download you can't skip is nvidia/diffusiongemma-26B-A4B-it-NVFP4 into ComfyUI/models/LLM/. The full LTX example workflow additionally expects ComfyUI-LTXVideo, comfyui-videohelpersuite, comfyui_controlnet_aux, comfyui_essentials, RES4LYF, and friends installed as custom-node dependencies.

Common issues

• Don't connect the H3 Reference Context here. That node declares H3 Ref2VA asset semantics and the LTX-2.5 readiness gate intentionally rejects it. Use the ordinary Context Hub for LTX.
• "Blocked on a cut" when you're sure the action fits. Director checks the generated opening against verified first-frame facts on I2V; a conditioning first frame fixes most of these by supplying the geometry Director would otherwise guess.
• Overloaded prompts for long videos. If you keep hitting validation on multi-action shots, remember shot/action density is budgeted against target_duration_seconds - raise the duration or thin the brief, don't fight the gate.