## TL;DR - [hellorob's outpainting comparison](https://x.com/hellorob/status/2057927569547440460) said Seedance 2.0 beat Gemini Omni on video outpainting, while LTX 2.3 came close at lower cost. - [CuriousRefuge's continuity workflow](https://x.com/CuriousRefuge/status/2057920807389806699) highlighted a new Stitcher tool built around Seedance 2 clips, aimed at hiding the join between continuation shots. - [NahFlo2n's ad workflow](https://x.com/NahFlo2n/status/2057732788187525391) showed a practical use case for reference-based ad variants, keeping one winning look while swapping outfit, product, creator angle, or offer. - Seedance 2.0 is already showing up as infrastructure, with [LumaLabsAI's launch post](https://x.com/LumaLabsAI/status/2057864318910161036) putting it inside Luma Agents and [egeberkina's Flova thread](https://x.com/egeberkina/status/2057178391225499843) using it inside a multi-step filmmaking agent.
You can open [Luma Agents](http://lumalabs.ai/app), browse a creator's [Mitte workflow](https://mitte.ai/?ref=on4ff72x), and even see Seedance used as the base model inside agent products like Flova in [egeberkina's setup post](https://x.com/egeberkina/status/2057178455566135647). The weirdly useful pattern across the feed is that Seedance keeps surfacing underneath other products, while hands-on tests like [hellorob's outpainting run](https://x.com/hellorob/status/2057927569547440460) and [CuriousRefuge's acting comparison](https://x.com/CuriousRefuge/status/2057610320198439186) frame it less as a novelty generator and more as a dependable motion layer.
[[tweet:https://x.com/i/status/2057927569547440460]]
The cleanest head-to-head in this evidence set is outpainting. [hellorob](https://x.com/hellorob/status/2057927569547440460) said Omni failed across multiple reference videos and prompts, while Seedance 2.0 "nailed it" and LTX 2.3 also performed well for much less money.
That lines up with a broader sentiment in [CuriousRefuge's Omni comparison](https://x.com/CuriousRefuge/status/2057929340562907451), which called Omni interesting for big action scenes and VFX-style effects but still less photorealistic than Seedance 2.0. [CuriousRefuge's acting comparison](https://x.com/CuriousRefuge/status/2057610320198439186) made a similar distinction on performance, saying Seedance felt more emotionally nuanced while Kling 3.0 often read broader and more theatrical.
[[tweet:https://x.com/i/status/2057920807389806699]]
Continuity is where the model starts looking genuinely useful for production, not just prompt demos. [CuriousRefuge](https://x.com/CuriousRefuge/status/2057920807389806699) described a workflow where Omni generates the next shot from the last frame of an existing clip, then a Seedance 2 Stitcher blends the seam so the handoff is harder to spot.
The mechanics are simple:
- Take the last frame from clip one. - Prompt a second model to generate what happens next from that frame. - Blend the two clips with a Seedance-focused stitcher to hide exposure and lighting drift.
A separate restoration test from [CuriousRefuge's archival comparison](https://x.com/CuriousRefuge/status/2057136653223796768) pushed the same idea in a different direction. There, Seedance reduced shakiness and distortion while staying closer to the source texture than Topaz Astra 2, which they said could drift too far into polish.
[[tweet:https://x.com/i/status/2057732788187525391]]
[[tweet:https://x.com/i/status/2057945638352867549]]
The most concrete commercial workflow here is not filmmaking, it is ad iteration. In [NahFlo2n's post](https://x.com/NahFlo2n/status/2057732788187525391), the claim is that one working aesthetic can become a reusable shell for new outfit, product, creator-angle, and offer variants without rebuilding the visual language from scratch.
Luma is pushing the same direction from the product side. [LumaLabsAI's testimonial graphic demo](https://x.com/LumaLabsAI/status/2057945638352867549), [LumaLabsAI's event graphics demo](https://x.com/LumaLabsAI/status/2057456303530807657), and [LumaLabsAI's sales graphics demo](https://x.com/LumaLabsAI/status/2057147807367102727) all frame Seedance inside Luma Agents as a system for turning one campaign brief into many channel-specific assets.
That makes the model look strongest when the task is constrained by an existing reference, not when the prompt is a blank page.
[[tweet:https://x.com/i/status/2057401313764196746]]
[[tweet:https://x.com/i/status/2057479267110314190]]
Creators keep getting the best results by feeding Seedance a lot more structure than a single text prompt. Across the strongest threads, the repeatable pattern looks like this:
1. Build source images or sheets in another model, often GPT Image 2 or Midjourney, as in [MayorKingAI's character sheet prompt](https://x.com/MayorKingAI/status/2057936981674635632) and [techhalla's storyboard workflow](https://x.com/techhalla/status/2057479267110314190). 2. Turn those images into a storyboard, sequence, or timing plan before animation, which [CuriousRefuge's previz workflow](https://x.com/CuriousRefuge/status/2057610320198439186) said works well for pacing and composition tests. 3. Tell Seedance exactly what must stay fixed. In [Artedeingenio's page-turn prompt](https://x.com/Artedeingenio/status/2057401317740429758), that meant preserving typography while only animating the illustrations. 4. Extend the sequence in chunks. [CharaspowerAI's continuation prompt](https://x.com/CharaspowerAI/status/2057499543470739542) split an anime fight into a first 15 seconds and a follow-up 15 seconds, instead of asking for the whole short at once.
The model still looks prompt-sensitive, but the prompt engineering is shifting upward into storyboards, character sheets, and timing scaffolds.
[[tweet:https://x.com/i/status/2057864318910161036]]
[[tweet:https://x.com/i/status/2057178391225499843]]
Seedance 2.0 is increasingly appearing as a source model inside larger creative stacks. [LumaLabsAI](https://x.com/LumaLabsAI/status/2057864318910161036) said it is live in Luma Agents, while posts from creators show it running through Mitte, Hailuo, Runway, Leonardo, Magnific, OpenArt, and Flova.
The most detailed stack description comes from [egeberkina's workflow breakdown](https://x.com/egeberkina/status/2057178525443186930), which says Flova's Story Driven Video skill split a project into:
- Spec - Storyboard - Character references - Prop references - Shot videos - Audio - Final assembly
Then [egeberkina's asset-generation post](https://x.com/egeberkina/status/2057178546167222733) said the agent built consistency assets before generating five shots with Seedance 2.0. That is new information beyond the comparison tests: the model is not only being judged as a standalone generator, it is being absorbed as the motion engine inside products that handle planning, review, and assembly around it.