11:34
AI video generation
Seedance 2.0, launched exclusively on Higgsfield, can convincingly execute the five current highest-virality short-video formats — transformations, first-person power "orbs," POVs, choreographed fights, and animations — and the video argues the real unlock is a free Claude "skill" that turns a plain-English idea into the long, detailed prompt the model needs.
Seedance 2.0 is presented as "the most anticipated model of the year," available exclusively on Higgsfield.
The presenter stress-tests it against five of the most viral short-form video formats currently circulating: transformations, orbs, POVs, fights, and animations, giving generations, exact prompts, and pro tips for each.
A free Claude "skill" was built to convert plain-English ideas into long, cinematic prompts automatically, since most viral-quality prompts are long and complicated and "most people don't write that way."
Format one, Transformations, follows what the video calls an "abnormal three-act structure": normal, chaos, back to normal (e.g., a girl eating a burger, a monster attack, then the girl eating unbothered again).
For transformations, the presenter runs an image-to-video vs. text-to-video side-by-side on the same scene: image-to-video uses separately generated Soul Cinema character/location/monster assets, while text-to-video uses none.
The video's stated tradeoff: both approaches work for a one-off viral clip, but image-to-video is "definitely worth the extra steps" if you want a recurring, consistent character.
Pro tip for realism: if generated skin/texture looks "too smooth and plasticky," adding "no 3D, no cartoon, no VFX" to the prompt is claimed to force Seedance into ultrarealism.
The video claims Seedance adds genre-appropriate details unprompted, e.g., inventing a visual gag in the background of a comedy scene without being asked.
Format two, Orbs, is a first-person format where the viewer's hands "crackle with power"; the video claims Seedance 2.0 is "one of the only models" that holds a consistent first-person perspective through chaotic action, with hands staying visible, reacting correctly to the environment, and correct scale.
In the orb examples, only the location image was supplied; the rest of the scene (powers, monster, action) came entirely from the text prompt.
Format three, POVs, is called "the hardest format to get right" because most models let the camera "drift" during fast movement/impact, breaking the illusion; the video claims Seedance 2.0 does not drift.
Technique for POVs: describing what the camera physically feels like at each beat, and keeping some POV prompts intentionally short ("single frame POV").
Pro tip for POVs: be explicit about what the camera is NOT doing — "no cuts, no zooms, and maybe natural head movement" — which the video claims is what keeps the perspective locked; without it, Seedance defaults to cutting between angles.
Format four, Fights, is claimed to generate not just hits but style, tension, and choreography, with camera work that "reads like someone actually shot this."
One fight example used a deliberately very long, detailed prompt for a six-shot action sequence on a train, all generated in 15 seconds in a single generation.
A second fight example used a minimal prompt ("I want her to kick his ass") with no location, only actors, to show a comedic scale-contrast (a small character vs. an enormous fighter) that the model committed to fully.
Format five, Animations, is framed as the most surprising: one example used a start frame and prompt for a dynamic action sequence, described as "production-level animation" made in about 5 minutes.
The second animation example used an extremely minimal prompt — "fight over 3D person with a TV person" — with the video claiming Seedance invented the fight choreography, style, characters, location, and slow-motion impact entirely on its own.
Stated overall workflow: drop the idea into Claude (which refines/enhances the prompt via the skill), refine if necessary, then generate on Higgsfield — described as "the easiest way to go viral right now."
Prompts and reference images used in the video are stated to be shared in the video description for viewers to reuse.
Transformations format — A viral video format built on an "abnormal three-act structure" — normal, chaos, back to normal — where a character faces a sudden threat and then returns to their mundane activity unbothered. Apply: Prompt a mundane action, interrupt it with a transformation/monster attack, resolve it, then have the character return to the original mundane action as if nothing happened (e.g., the pink-haired girl who becomes a monster, eats an attacker, turns back, and resumes eating her burger).
Orbs format — A first-person format where the viewer's own hands are shown "crackling with power," relying on the model holding a consistent first-person perspective and correct hand/environment interaction through chaotic action. Apply: Supply only a location reference image and describe the power-fantasy action in text (e.g., electric powers in a post-apocalyptic city defeating a colossal monster with an electrical beam), optionally specifying a lens style like fisheye.
POVs format — The hardest of the five formats: a first-person camera that must commit to and hold one perspective through fast movement and impact without "drifting" into cut angles. Apply: Describe what the camera physically feels like at each beat, keep some prompts intentionally short and single-frame-POV focused, and add explicit negative camera instructions (no cuts, no zooms, only natural head movement) to keep the perspective locked.
Fights format — A format where the model is claimed to generate not just impacts but full choreography, tension, and camera work that reads like it was actually filmed. Apply: For maximum choreography control, write a very long, detailed multi-shot prompt (e.g., a six-shot fight sequence on a train); for a shorter, contrast-driven beat, give only the actors and a minimal instruction (e.g., "I want her to kick his ass") and let the model handle staging.
Animations format — An anime/animation-style action format in which the model is claimed to invent fight choreography, character style, location, and slow-motion impact largely on its own from minimal input. Apply: Optionally supply a start frame, then write an extremely short prompt (the video's example was "fight over 3D person with a TV person") and let the model autonomously fill in the rest of the scene.
Claude prompt-enhancement skill — A free custom Claude "skill" built by the presenter that converts a plain-English scene idea into a long, detailed, cinematic Seedance prompt. Apply: Type a plain-English description of the desired scene into Claude running the skill (e.g., "Enhance this prompt. The pink-haired girl turns into a monster..."), then take the resulting cinematic prompt and paste it into Seedance/Higgsfield.
Image-to-video vs. text-to-video workflow — Two generation paths for the same scene: image-to-video, which feeds Seedance separately generated character/location/monster images, versus pure text-to-video with no reference images. Apply: Use text-to-video for a one-off viral clip; use image-to-video (generating character/location/asset images in Soul Cinema first, then animating them) when a recurring, visually consistent character is needed across multiple videos.
Soul Cinema asset generation — A tool used in the video to generate the still character, location, and monster images that are then fed into Seedance as image-to-video inputs. Apply: Generate each scene element (main character, secondary character, location, monster) as separate still images in Soul Cinema before animating them together in Seedance.
"No 3D, no cartoon, no VFX" realism prompt — A stated prompt add-on to force Seedance into ultrarealism when generated skin/texture looks too smooth or plasticky, particularly for monsters. Apply: Append the phrase "no 3D, no cartoon, no VFX" to a prompt whenever a generated creature or character reads as artificially smooth or plasticky.
POV camera-lock instruction — A specific negative-instruction technique for keeping a POV shot from breaking into cuts or zooms by telling the model explicitly what the camera is not doing. Apply: Add lines like "no cuts, no zooms, natural head movement" to POV prompts to prevent the model from defaulting to cutting between angles and breaking the first-person illusion.
Across all five formats, the actual prompting lever emphasized isn't raw prompt length but explicit control of camera behavior — what the camera does and, just as importantly, does NOT do (no cuts, no zooms) — suggesting perspective-lock instructions matter more than descriptive detail for POV-style shots.
The video treats the transformation and orb formats as driven by the same underlying mechanic: a tension-then-release loop that makes a clip rewatchable, which it links directly to how well a clip performs with "social media algorithms" rather than to visual spectacle alone.
The image-to-video vs. text-to-video comparison functions as an implicit cost/benefit demonstration: asset-based generation (extra Soul Cinema step) buys character consistency across multiple videos, while text-to-video is framed as sufficient only for single, disposable clips.
The video's closing example is deliberately the most minimal prompt of the entire video ("fight over 3D person with a TV person") placed right after the most maximal prompt (the long six-shot train fight), using that contrast to argue the model's autonomous scene invention is the real selling point rather than prompt-writing skill.
Claude is positioned not as a side tool but as an integrated stage of the pipeline — the video frames plain-English input into Claude, not the raw model prompt syntax, as the actual user-facing interface for Seedance.
«These videos keep blowing up. Millions of views with no cameras and no crews involved.»
— 00:00
«It follows an abnormal three-act structure. Normal, chaos, back to normal.»
— 01:17
«if the skin looks too smooth and plasticky, add no 3D, no cartoon, no VFX to the prompt. This forces Seederns into ultrarealism.»
— 02:57
«POV is the hardest format to get right.»
— 06:20
«A little pro tip for POVs, be explicit about what the camera is not doing. No cuts, no zooms, and maybe natural head movement.»
— 07:28
«The camera treats her as the dominant force from the very first frame.»
— 09:23
«Nope. Brace yourself. The prompt was fight over 3D person with a TV person. That's it.»
— 10:40
«Drop in the idea into Claude, refine if necessary, and generate on Kickfield. This is the easiest way to go viral right now.»
— 11:09
Reception
Strong technical interest in Seedance 2.0 itself, but overshadowed by substantive complaints about billing practices, service reliability, broken support, and unresolved technical bugs.
This is a promotional capability showcase and tutorial for Higgsfield's Seedance 2.0 (bundled with the presenter's own free Claude prompt-enhancement skill), built entirely from curated best-case generations across five format categories, with concrete prompting techniques (three-act structure, negative camera instructions, image-vs-text-to-video tradeoffs) but no disclosed failure rate or selection criteria for the examples shown.
11:34