Shot failure diagnosis: locate the problem in 60 seconds
AI video generation failure comes in only two kinds: it doesn't run (queuing, errors, stuck, interrupted) and it runs but is wrong (corrupted frames, face drift, not moving as prompted). The root cause sits in different layers—fix the wrong layer, and a hundred rerolls are just wasted credits.
This page is the triage process designed for this course: locate the layer in 60 seconds first, then match your case. The specific original error messages for local ComfyUI are not covered on this page—go to the troubleshooting manual in the comfyui section and match by the original text; every recommendation is tagged with its applicable environment (cloud / local / universal).
What you'll be able to do after learning
- Split a failure into two layers within 60 seconds: it doesn't run (environment layer) or it runs but is wrong (content layer), and state what to check first
- State the troubleshooting starting points for cloud and local: on cloud, check task status; locally, check the original error messages in the ComfyUI server console
- Triage the three types of content problems—corrupted frames, wrong character, mismatched action—against the table, knowing whether to change the prompt, swap the first frame, or change the shot size
- When local VRAM is insufficient, follow the downgrade path in order: pick the right template → reduce size and frame count → system parameters → community quantized version (marked "community resource")
- Know when to stop: retry cap of 3 attempts per shot, then hand over to the 4.5 downgrade path
Your current predicament
- "I clicked generate and image-to-video is stuck halfway—I don't know if the cloud is queuing, local VRAM blew up, or it really crashed"
- "The video did come out, but Lao Chen's face is a blurry mess and his hand is fused to the cake box—I don't know whether to change the prompt or swap the image"
- "The prompt says he falls into a puddle, but he just spins in place; I rerolled three times and he spun all three—only then did I realize it's not a luck problem"
When to use this technique
When you need to:
- Locate before acting after a failure, turning "why it failed" into "which variable to change"
- Spend cloud credits and local electricity where they count—don't leave problems doomed to fail to the reroll lottery
And you don't want to:
- Change the prompt, first frame, and size all at once, so even if it works you don't know which one did it
- Only discover on attempt 8 that this shot simply cannot be generated
🎒 Preparation before starting
- Completed 4.1 your first shot: having seen one success, you'll know which link a failure gets stuck on
- Have the 2.4 storyboard and three fixed text sections on hand: 3.1 character card, 3.3 scene card, 3.4 style card—diagnosing content problems requires checking against them
- Can state your environment: cloud (such as Comfy Cloud) or local ComfyUI—if unsure, go back to 1.3 the two routes
- Keep every generated file, with version numbers per the 3.5 naming convention—diagnosis relies on comparing versions
Core idea
Split a failure into two layers first; different layers call for completely different actions (the teaching process designed for this course):
| Layer | Typical symptoms | First step |
|---|---|---|
| Environment layer: it doesn't run | Queuing, failure, interruption, frozen, nodes showing red | First determine cloud or local, then check the "it doesn't run" group |
| Content layer: it runs but is wrong | Corrupted frames, wrong character, mismatched action, first and last frames not matching | Match against the "it runs but is wrong" triage table, change one variable and retry |
Retrying is useless for environment-layer problems—if VRAM isn't enough, it isn't enough; blind rerolling is also useless for content-layer problems—the root cause in the prompt grows the same fruit every time.
The three laws of diagnosis (designed for this course):
- Change only one variable at a time: prompt, first frame, size, workflow—touch one thing per attempt
- Archive before changing: keep failed versions per the 3.5 naming convention; only comparison tells whether the change worked
- Count retries, cap at 3 attempts: once the limit is reached, hand over to 4.5 retry and replacement strategy, don't fight on
First troubleshooting pass: 60-second quick check
Ask yourself the five questions and locate the layer and group within a minute (the triage question list designed for this course):
# 1. Did the video come out?—No → environment layer, continue with 2, 3; came out but wrong → content layer, go straight to 4
# 2. Cloud or local?—Cloud: check the platform's task status and error messages; local: go through the ComfyUI server console to find the original error text
# 3. Which template are you running locally?—"Wan2.2 5B" (official claim: runs on 8GB of VRAM) or
# "Wan2.2 14B I2V" / "Wan2.2 14B FLF2V" (official repo script, single-GPU claim: 80GB)—if it won't run, lower resolution and frame count first
# 4. What type of content problem?—Broken visuals / wrong character / motion not matching; for first-last-frame shots, also check the two end frames
# 5. How many retries for this shot?—Cap at 3; beyond that, stop and go to 4.5After asking question 1, you'll know which group to go to.
Common Issue Troubleshooting
It won't generate: generation failure, hanging, and interruption (AI video generation failure)
Cloud: tasks queue for a long time, or fail outright [Environment: cloud]
Cause: cloud tasks must pass through several gates—queue, platform resources, template and account status; specific errors vary by platform, and this course has not verified each cloud platform's error documentation.
Solution:
- Refresh the task list or re-enter the workflow page to confirm whether the task is queued, running, or failed
- Drop down one tier of specs and resubmit: smaller resolution, shorter duration (in ComfyUI workflows this is the frame count length)
- Stop after two consecutive failures: consult the platform's official documentation or contact platform support, and include the original error message
Local: interrupted halfway through, suspected VRAM shortage [Environment: local]
Cause: insufficient VRAM is the most common local generation failure. There are two official claims, corresponding to different ways of running: ComfyUI running 5B—the official tutorial says that with native offloading, 8GB of VRAM is enough (docs.comfy.org Wan2.2 tutorial); running with the official repo script, the same 5B needs at least 24GB on a single GPU (e.g., RTX 4090), and 14B-series single-GPU commands are labeled 80GB (GitHub Wan-Video/Wan2.2).
Solution (step down in order):
- Identify the template: confirm which one you're running—search for "Wan2.2 5B" or "Wan2.2 14B I2V" / "Wan2.2 14B FLF2V"; if even 5B won't run, close other VRAM-consuming programs and try again
- Lower the specs: 14B failing to run is normal—the official FLF2V template defaults to a smaller resolution (designed for low VRAM); drop resolution and frame count by one tier each. Reference for vertical 720P: the TI2V-5B README lists 704*1280—run the defaults first, then scale up once it works
- System-level fix: ComfyUI startup-parameter-level VRAM reduction methods, see ComfyUI VRAM shortage troubleshooting
- Still not enough: switch to community GGUF quantized versions (bullerwins, QuantStack repos + City96/ComfyUI-GGUF nodes, marked "community resources"), see ComfyUI GGUF quantization
💡 Confused by "runs on 8GB" vs. "24GB to start"?
No conflict—these are the claims for the same model under two ways of running: ComfyUI native offloading is one; the official repo script (generate.py) is the other—the 5B single-GPU 24GB command itself carries the three VRAM-reducing parameters --offload_model True --convert_model_dtype --t5_cpu, and the official advice for OOM (Out-of-Memory) is also to add these three. Match your case accordingly; don't hold B to A's claim.
Local: generation so slow you suspect it's frozen [Environment: local]
Cause: video generation is inherently slow—compare against the official claim before deciding whether to intervene: with TI2V-5B without dedicated optimizations, a single consumer-grade GPU generates 5-second 720P video within 9 minutes (GitHub Wan-Video/Wan2.2 speed statement).
Solution:
- Use the official claim as your yardstick: 5B running 5-second 720P within 9 minutes is normal—don't kill the process midway
- Far beyond that magnitude: it's not a matter of waiting—go back to the previous item and lower the specs (drop resolution and frame count by one tier each, then retry)
- Progress completely stuck, no console output: note the symptoms and take the original error text to ComfyUI troubleshooting overview
Workflow nodes missing or showing red [Environment: local / cloud]
Cause: the official tutorial attributes missing core nodes to two things: ComfyUI is not the latest version (Nightly); or node import failures at startup. Cloud has one more: Comfy Cloud only follows updates after the ComfyUI stable release, so new core nodes can't be found before they reach the stable version (docs.comfy.org Wan2.2 tutorial).
Solution:
- Local: update ComfyUI to the latest version; after restarting, check the server console for node import failure messages
- Cloud: confirm it's not a new node that hasn't reached the stable version—wait for the next stable release
- Red nodes, specific model-loading errors: follow the original error text to ComfyUI troubleshooting overview; nodes and models each have dedicated pages
It generates but comes out wrong: broken results, wrong characters, motion not matching
The video came out but is unusable—first classify it against the table (the triage table is designed for this course; content problems are unrelated to cloud vs. local, so all items below are marked [Environment: universal]):
| Symptom | One-line cause | Where to make the first cut |
|---|---|---|
| Broken visuals: twisted limbs, smeared faces, objects fused together | Asking the model to do too much at once, or the subject is too small | Subtract from the motion |
| Wrong character: face drifts, looks like a different person | The anchor isn't pinned: the first frame image was wrong to begin with | Check the image first, then the prompt |
| Motion not matching: doesn't follow the prompt, improvises on its own | Verbs not concrete, or the prompt conflicts with the first frame | Change the verbs, split the task |
Broken visuals: twisted limbs, faces mushed into a blur【Environment: General】
Cause: In most cases the model isn't broken—the task is just too hard: too many actions crammed into one shot, or the subject is too small in the frame. 06 Old Chen's full-shot fall is a disaster zone: small subject, big motion, lots of rain-streak interference.
Solution:
- Subtract from the single-shot task line: write only one continuous motion ("foot slips, body pitches forward, falls sitting into a puddle") and delete all the extra theatrics—spinning, head turns, slow motion. Pick verbs from 4.2
- First self-check the first-frame image: if the image itself is broken (six fingers, mushed face), go back to 3.2 reference images and regenerate it—editing images is free, don't burn generations
- If the subject is too small, change the shot scale: when the full shot breaks the face, switch to a medium or close shot and regenerate; leave the full-shot information to adjacent shots
- Re-prepare reference images per "even lighting, neutral expression" (the Runway Gen-4 help page's advice on reference-image inputs, borrowed by this course as a general standard)
Wrong character: the face drifts, looks like a different person【Environment: General】
Cause: The anchor wasn't pinned. The anchor for image-to-video is the first-frame image—if the person in the first frame isn't Old Chen, the output naturally won't be; or the first frame is right but the fixed description block wasn't pasted into the prompt, so the model improvised.
Solution:
- First check the first-frame image: compare it item by item against the fixed description block in the 3.1 character card—square face, short black hair, yellow rider vest; if the image is wrong, go back to 3.2 and regenerate
- Then check the prompt: paste the entire fixed description block as-is (3.4 four-part formula), don't just write "a delivery rider"
- For face drift across shots, rescue with continuation frames: use a screenshot of the previous shot's final frame as this shot's first frame (3.2's "continue from previous shot's final frame"), carrying the character's state across the cut point
- Set the bottom line first: this course does not promise absolute character consistency—reference images plus fixed descriptions are a mitigation, not a guarantee
Motion mismatch: doesn't move as prompted, or adds its own theatrics【Environment: General】
Cause: The verb isn't concrete ("emotional breakdown" can't be drawn), three things crammed into one shot, or the prompt fights the first frame—the first frame shows him standing, the prompt says running.
Solution:
- Replace verbs with filmable ones: write only one thing at a time, with complete subject-verb-object (4.2's verb usage)
- Cross-examine the prompt against the first frame: the first frame is the starting point; the prompt should describe motion departing from that starting point, don't make the model first imagine how to get there
- Borrow the official prompt-extension mechanism: the Wan2.2 official repository recommends expanding prompts before generation—the script's switch is --use_prompt_extend; the extension model for T2V uses Dashscope's qwen-plus, I2V uses qwen-vl-max, or a local Qwen (--prompt_extend_target_lang 'zh' can specify Chinese); the official ComfyUI tutorial's workflow doesn't mention this mechanism, but the idea can be borrowed—before generating, use any LLM to expand the single-shot task line into a complete prompt, then paste it into the workflow
- After fixing, retry with the count still running: capped at 3 attempts
First-last frame shots: the two ends don't match the uploaded images【Environment: General】
Cause: In the first-last frame workflow, the two ends are the two images you upload—swap them and the shot runs backwards; if the frame images aren't from the same source, a smooth middle is useless.
Fix: Go back to 4.3's acceptance triple-check: check the clip head against the first frame and the clip tail against the last frame, once each; only if both ends match but the middle is odd is it a prompt issue. When uploading images, trust the node labels, not the example images.
Still not working?
- Archive first, then ask: keep the failed versions per the 3.5 naming convention, and log this shot into the 1.4 cost sheet—symptom, what you changed, result; one line is enough
- Local users: take the complete original error text to the ComfyUI troubleshooting overview and match it by the original text; for VRAM issues go to the insufficient VRAM page; if it still fails, take the full checklist below to the ComfyUI official repository Issues
- Cloud users: this course has not verified the error documentation of each cloud platform—consult your platform's documentation or contact platform support; don't force-fit local ComfyUI solutions onto the cloud
- Content problem keeps failing: this shot may simply be a difficult birth—go to 4.5 retry and replacement strategies, which has downgrade routes and criteria for "letting it go"
Prepare five things before asking (so anyone can reproduce your problem):
| What to bring | Example |
|---|---|
| Environment and hardware | Cloud: platform name + template name; local: GPU model and VRAM |
| Workflow | Template library search name, e.g. "Wan2.2 14B FLF2V" |
| Original error text | Local: copy the entire server console output; cloud: screenshot of the platform error message |
| Full prompt text | Paste the whole thing, not just the single-shot task line |
| What you've already tried | Which variable you changed, how many retries, each result |
Checkpoint ✅
- 60-second triage you can state: won't run means environment layer, runs but wrong means content layer, and what to check for each
- You can state the starting point for cloud vs. local: task status / original server console error
- You know the first cut for each of the three content problems: subtract for broken visuals, check the image first for wrong character, swap verbs for motion mismatch
- You can recite the local VRAM downgrade route: pick the right template → reduce size and frame count → system parameters → community quantized versions (marked "community resource")
- Follow the three rules of diagnosis: one variable at a time, archive before changing, capped at 3 attempts
FAQ
What should I do if image-to-video reports an error?
First identify the environment before taking action: in the cloud, refresh the task list or re-enter the workflow page to confirm whether the task is queued, running, or failed; locally, check the ComfyUI server console for the original error message. Insufficient local VRAM is the most common cause of failure—downgrade in the order of "pick a template → reduce size and frame count → adjust system parameters → use a community quantized version"; see the "It won't run" section. Take the original error message to ComfyUI Troubleshooting overview and match it against the original text.
What should I do if image-to-video fails to generate content?
If the video doesn't come out, it's an environment-level issue—retrying won't help; not enough VRAM is just not enough. First do the "60-second quick check": confirm whether you're on the cloud or local, and which template you're running ("Wan2.2 5B" or "Wan2.2 14B I2V" / "Wan2.2 14B FLF2V"), then drop the resolution and frame count down one level each and retry. Cap retries at 3 per shot; beyond that, hand it over to the 4.5 retry and replacement strategy.
What to do when a shot generated by AI storyboarding errors out?
If a node shows red or is missing, check two things first: your ComfyUI version isn't the latest (Nightly), or some nodes failed to import at startup—update ComfyUI locally to the latest version, then restart and check the console messages. On the cloud (Comfy Cloud), you have to wait for ComfyUI to release a stable version before the update follows; new core nodes won't be found until they make it into a stable release, so just wait for the next stable version.
Lesson Summary
✅ This page on shot-failure diagnosis turns AI video generation failures from mysticism into a process:
- Layer first: won't run means environment layer (cloud checks task status, local checks console and VRAM); runs but wrong means content layer (three-way triage)
- Environment layer checked per environment: cloud failures end with "consult platform documentation / contact support"; local VRAM shortage follows the downgrade route—don't mix the two VRAM standards
- Three cuts for the content layer: subtract for broken visuals, check the first-frame image first for wrong character, swap in concrete verbs for motion mismatch; first-last frame shots additionally check the two ends
- Three rules of diagnosis: one variable at a time, archive before changing (3.5), retry capped at 3 attempts (4.5)
Key standards recap:
| Item | Standard | Source |
|---|---|---|
| ComfyUI running 5B | with native offloading, 8GB of VRAM is enough | ComfyUI official tutorial |
| running with the official repository scripts, 5B | requires at least 24GB on a single card (e.g., RTX 4090) | Wan2.2 official README |
| 14B series | the single-card command notes 80GB | Wan2.2 official README |
| 5B speed benchmark | with no special optimization, a single consumer-grade GPU produces 5 seconds of 720P video within 9 minutes | Wan2.2 official README |
| low-VRAM alternative | community GGUF quantized versions (bullerwins / QuantStack + City96 nodes) | the "Community Resources" section of the official tutorial, labeled "Community Resources" |
Next Lesson Preview
Now that we've found the root cause, the next lesson tackles "what to do next": 4.5 Shot Retry and Replacement Strategy: at most 3 retries per shot.
You will learn:
- How to calculate the retry limit: 3 retries per shot, and where the counting starts
- How to follow the downgrade path: change the prompt, change the first frame, change the workflow, or let it go
- How to record failure costs: backfill the 1.4 cost table, so the next film pays less tuition
Related Tutorials
- ComfyUI Troubleshooting Overview—match local errors to the entries by their original text; separate pages cover nodes, models, and VRAM
- ComfyUI Insufficient VRAM and Stuttering—systematic solutions for OOM: VRAM-reduction measures at the launch-parameter level
- Shot Retry and Replacement Strategy—the next step after diagnosis: a limit of 3 retries per shot, downgrade paths, and cost backfilling