Seedance2

Guide

Seedance 2.0 Troubleshooting — Fix Common Generation Problems

Most weak outputs come from a small set of repeatable issues: unclear references, overloaded prompts, mismatched motion expectations, or pushing too many changes into one generation. Use this page as a debug checklist before you blame the model.

Last updated: Last verified:

Source basis and reading boundary

These guides are written as third-party reference summaries, not official product documentation or support content.

Source basis

Fix character drift before changing everything else

If identity changes across shots, simplify the setup first. Use fewer references, remove extra costume or lighting changes, and repeat the core identity phrase in each request. Character drift often happens when a workflow mixes one hero character with too many environment or style changes at the same time. Stabilize the face and outfit first, then add camera moves and secondary objects later.

Reduce prompt overload and conflicting instructions

Ignored prompts often come from too many instructions competing inside the same paragraph. Put the critical subject, action, and camera cue first. Trim decorative adjectives that do not affect the shot. If one prompt asks for fast motion, precise lip-sync, cinematic shallow depth of field, multiple props, and a sudden style shift all at once, some details will get dropped. Break ambitious ideas into smaller passes or shorter shots.

Debug motion, timing, and extension step by step

Unnatural motion and bad extension results often improve when you shorten the clip and simplify the action. Use phrases like slow, controlled, continuous, or steady when you need readable motion. If you are extending a clip, protect the last good frames and ask for a natural continuation rather than a new event. For lip-sync and native audio, keep dialogue short first and check that the scene framing still supports mouth visibility.

Negative prompts do not work

Seedance 2.0 does not have a negative prompt field, and phrasing like 'no blur,' 'don't show hands,' or 'avoid text on screen' can actually increase the presence of the unwanted element. The model parses the descriptive noun and treats it as content to include. The fix is always positive phrasing: replace 'no blur' with 'sharp, in-focus, high clarity.' Replace 'don't show hands' with a framing instruction that naturally excludes them, such as 'close-up on face, shoulders up.' This single habit change eliminates a large category of frustrating results.

Generation failed error

A 'generation failed' message usually comes from one of four causes: the uploaded file format is not supported (stick to JPG, PNG, WebP for images and MP4 for video), the file exceeds size limits (roughly 30 MB for images and 50 MB for video), the content triggered the platform safety filter, or the queue is congested and timed out. Check file specs first — format and size issues account for the majority of generation failures. If the file is valid, review the prompt for content that might trip the safety filter. Queue congestion failures usually resolve by retrying after a few minutes.

Prompts getting flagged by safety filter

The safety filter is more likely to flag vague context than specific professional language. A prompt like 'person on operating table' is more ambiguous and risky than 'medical training video: surgeon performing routine procedure in well-lit operating room, educational context.' Be explicit about the professional setting, intended use, and tone. Avoid shorthand that could be misinterpreted. If a prompt is repeatedly flagged, rephrase the action with more surrounding context rather than removing descriptive words.

Hands and fingers look distorted

Hand distortion is a known limitation across current video generation models, including Seedance 2.0. To minimize it: use slow, deliberate gestures instead of fast or complex hand movements; make the hand action explicit in the prompt ('gently picking up a cup with right hand, fingers wrapped around the handle'); avoid poses where hands overlap or are partially hidden behind objects; and keep the hand action as the primary focus of the shot rather than a secondary detail. If hands are not essential to the scene, frame the shot to exclude them.

Text and logos become garbled

AI video models struggle with rendering readable text, numbers, and detailed logos. If your scene requires on-screen text, keep it very large, simple, and limited to 2–3 characters. For anything more complex — brand logos, subtitles, UI elements, or paragraphs — plan to add them in post-production as an overlay. This is a standard practice in professional AI video workflows and is not a Seedance-specific limitation. Baking text into the generation prompt almost always produces illegible or warped results.

Too many characters cause warping

Scenes with more than two characters frequently produce identity blending, body warping, or spatial confusion. Limit each shot to two characters maximum with clear spatial separation — for example, 'woman on the left side of frame, man on the right, one meter apart.' Describe each character's position, pose, and distinguishing features explicitly. For group scenes, consider generating characters separately and compositing in post, or use a wide shot where individual detail matters less.

Prompt too short vs too long

The practical sweet spot for Seedance 2.0 prompts is 30–100 words. Prompts under 20 words give the model too much freedom to fill in defaults, which produces unpredictable results. Prompts over 150 words cause instruction dropout, where the model quietly ignores details it cannot reconcile within one generation. If you have more than 100 words of instructions, split them across reference images, SCELA-structured sections, or multi-shot scripts rather than packing everything into a single paragraph.

Style chaos from multiple references

When you upload multiple reference images with different lighting conditions, color palettes, or art styles, the model tries to average them, producing a muddled result. The fix is to use one strong style anchor image that defines the overall look, and ensure any additional references share the same lighting direction, color temperature, and visual language. If you must mix references from different sources, pre-process them in an image editor to unify white balance and contrast before uploading.

Queue wait times during peak hours

Free-tier users can experience wait times of two hours or more during peak usage periods, typically weekday evenings and weekends in major time zones. Strategies to reduce wait time: generate during off-peak hours (early morning UTC), use the paid tier for priority queue access, batch your requests and submit them before peak hours, and avoid re-submitting the same request multiple times, which only adds to queue congestion. Paid tier users typically see generation times under 5 minutes.

Disabled or suspended features

As of the current Seedance 2.0 release, Human Reference Input and Face-to-Voice are suspended features. Do not include them in production workflows or expect them to be available. Check the official Dreamina and Seed project pages for reactivation announcements before building workflows that depend on these capabilities. Planning around suspended features is a common source of wasted production time.

Examples & sources

Debug pattern: strip the prompt back to one action

When a generation fails, remove style flourishes and secondary actions until one clean motion works. Then add detail back one layer at a time.

Same character as reference, medium shot, walking forward slowly, neutral expression, stable background, smooth handheld camera, no extra props.

Debug pattern: shorten before you extend

If a long clip breaks, first generate a shorter good clip, then extend from the last stable frame. This isolates continuity issues from prompt issues.

Generate a 4-second base shot first. After approval: continue from last frame, same angle, same speed, same rain reflections, preserve subject position.

Debug pattern: replace weak references before rewriting prompts

Many failed generations improve more from cleaner source material than from longer text. Check whether the input image already matches the final camera angle and lighting you want.

Debug pattern: flip negative phrasing to positive

If your output keeps showing the exact thing you wrote 'no' about, the model is reading the noun, not the negation. Rewrite the instruction as a positive description of what you do want.

Before (broken): 'no blur, no shaky camera, don't show background people.' After (fixed): 'sharp focus, locked-off tripod shot, empty background, single subject centered in frame.'

Debug pattern: isolate the failing variable

When a multi-reference generation fails, remove references one at a time and regenerate. The reference whose removal fixes the issue is the conflict source.

Test 1: character ref only, no style ref. Test 2: style ref only, no character ref. Test 3: both refs but with matched lighting. Compare all three to find the conflict.

Debug pattern: reduce character count for warping issues

If a multi-character scene produces body warping or identity blending, split the shot into single-character generations and composite in post-production.

Shot A: woman standing on left side of frame, solo, cream jacket, neutral pose. Shot B: man standing on right side of frame, solo, dark suit, same background and lighting. Composite both in editing software.

Frequently asked questions

Why does the character keep changing between shots?

The workflow is usually carrying too many changing variables. Start with fewer references, keep outfit and lighting stable, and repeat the same identity phrase in every shot before adding more complexity.

Why are important prompt details being ignored?

The most important instructions may be buried inside decorative text or conflict with each other. Move the subject, action, camera direction, and key constraints to the front of the prompt and cut anything that does not change the shot.

What should I check when motion looks distorted or too busy?

Reduce action complexity, shorten the shot, and choose one readable camera move. Fast movement, complex choreography, and multiple moving objects in one pass increase the chance of artifacts.

How can I improve weak lip-sync or awkward dialogue shots?

Keep the spoken line short, make sure the mouth is visible enough in frame, and avoid asking for heavy action at the same time. Public workflows usually test simple talking shots first before adding more movement.

What is the safest order for debugging a broken workflow?

Check access and mode first, then references, then prompt structure, then motion complexity, and only after that try extension or audio layers. This order prevents you from fixing three variables at once.

Why does writing 'no blur' actually make my output more blurry?

Seedance 2.0 does not support negative prompting. The model reads the noun 'blur' and treats it as a content signal. Replace negative phrasing with positive alternatives: 'sharp, in-focus, high clarity' instead of 'no blur,' or 'steady tripod shot' instead of 'no camera shake.'

What causes a 'generation failed' error and how do I fix it?

The most common causes are unsupported file formats, files exceeding size limits (30 MB for images, 50 MB for video), safety filter triggers, or queue timeouts. Check file format and size first, then review the prompt for ambiguous content that might trip the safety filter. Queue failures usually resolve by retrying after a few minutes.

How do I fix distorted hands and fingers in my generations?

Use slow, deliberate gestures and make the hand action explicit in the prompt. Avoid poses where hands overlap or hide behind objects. If hands are not critical to the scene, frame the shot to exclude them naturally. Hand distortion is a known limitation of current video generation models.

How many words should my prompt be?

The sweet spot is 30–100 words. Under 20 words gives the model too much freedom to fill in defaults. Over 150 words causes instruction dropout where details get quietly ignored. If you need more than 100 words, split the instructions across reference images, structured sections, or multi-shot scripts.

Why do my multi-character scenes look warped or blended?

Scenes with more than two characters frequently produce identity blending and spatial confusion. Limit each shot to two characters max with explicit spatial separation. For group scenes, generate characters separately and composite in post-production.

Are Human Reference Input and Face-to-Voice available in Seedance 2.0?

No. Both features are currently suspended. Do not include them in production workflows. Check the official Dreamina and Seed project pages for reactivation announcements before planning workflows that depend on these capabilities.

Related guides