A generation fails, and the instinct is to hit "run" again. Sometimes that works. Often it doesn't, because the failure wasn't random — it was a signal. In 墨川导演台 (MiniMax H3 AI Director's Console), that signal lives in two places: the AI logs and the system logs. Learning to read them turns guesswork into a five-minute fix.
Think of it like a film set. The AI log is the script supervisor's notes — what was supposed to happen in each shot. The system log is the stage manager's radio chatter — what the lights, cameras, and cloud rig were actually doing. When a take fails, you need both notebooks.
This walkthrough covers where to look, what to look for, and how to fix the most common failure patterns. Everything here is based on the console's real logging and deployment features.
---
H2: Where the Logs Live (and What Each One Tells You)
Before debugging anything, know your two windows.
AI logs track the creative pipeline: what the AI screenwriter returned, what prompt went to the image generator, which storyboard card was queued, what the voice synthesis step produced. If your script didn't land in the script library, or a storyboard card shows a failed status, this is your first stop.
System logs track the machine underneath: SSH connection state, cloud ComfyUI startup, tunnel port assignments, GPU instance health. If the console can't reach your cloud server, or ComfyUI never turned green, this is where the truth is.
A quick rule of thumb:
- *Creative output looks wrong or empty* → read AI logs first.
- *Nothing happens at all, or generation stalls at zero* → read system logs first.
- *Something worked yesterday and doesn't today* → read system logs, then AI logs.
Both are visible in real time inside the console, so keep them open in a second panel while you work. Debugging blind is like color-grading with the monitor off.
---
H2: A Step-by-Step Debugging Walkthrough
Here's the sequence I use whenever a generation fails. It moves from cheapest fix to most expensive, so you don't tear down a working cloud instance over a typo.
Step 1 — Confirm the failure type. Is it a hard error (the task stopped), an empty result (task "completed" but no video), or a timeout (task sits in queue)? The AI log will usually show the task's last state. Note the exact step that failed: script generation, first-frame image, video generation, or audio.
Step 2 — Check the system log for connection health. If the failure happened at the video step, glance at the system log for SSH drops, tunnel port conflicts, or ComfyUI restart messages. The console isolates local tunnel ports per user, so multi-user collisions shouldn't happen — but a cloud instance that was shut down or ran out of credit will absolutely show up here. If you see connection errors, jump to the cloud deployment section below before touching any prompts.
Step 3 — Verify the cloud ComfyUI is actually running. In the 云部署 (Cloud Deployment) page, confirm that "验证 SSH" passed and that "启动云端 ComfyUI" reached the green success state. If the green light is missing, no amount of prompt tweaking will help. Re-verify SSH, then restart ComfyUI and watch the system log during startup.
Step 4 — Read the AI log for the failing step's input. This is the most underrated habit. The AI log shows what the console *sent*. Nine times out of ten, a "failed generation" is a bad input: an empty prompt field, a storyboard card with no first-frame image attached, a reference image that never uploaded to the asset library, or a voice line with no matching character.
Step 5 — Isolate with a single-shot test. Use single-shot generation on the failing storyboard card instead of a batch run. If the single shot succeeds, the problem is in the batch queue or one specific card. If it fails identically, the problem is in that card's inputs or the model step itself. This one trick saves enormous time.
Step 6 — Fix, re-run, and log your own notes. Once it works, write down what changed. Your personal debug log becomes more valuable than any documentation.
---
H2: Common Failure Patterns and Their Fixes
These are the patterns you'll meet most often. Match the symptom to the section, then act.
"Script never appeared in the script library." The AI screenwriter replies in chat and writes drafts into the script library — but only when the conversation actually produces a draft. If you asked a vague question, you got a vague answer and no draft. Rephrase as an explicit request for a six-segment script and storyboard draft, in one of the five supported languages (Chinese, English, Spanish, Arabic, Japanese). Also check that your LLM API key is configured if you've connected your own.
"Image generation fails or returns nothing." Open the AI log and look at the prompt that was sent. Common culprits: a text-to-image or image-to-image request with a missing reference image (make sure it's uploaded in the asset library first), or a character-consistency request pointing at a character that doesn't exist yet. Re-upload the reference, re-select the asset, retry as a single shot.
"Video generation stalls or errors at the video step." Go to system logs. Confirm the cloud ComfyUI is green. If you're on 仙宫云 (Xiangong Cloud), remember the test flow requires the private mirror: register fresh at the official link with invite code B1H9M6, complete verification, send your account ID to customer service, and only then will the director's console image appear for deployment. A missing or stale instance is the single most common cause of "generation failed" at this step. If you're on AutoDL, Alibaba Cloud, Tencent Cloud, or a custom SSH server, verify host, port, and password in the cloud deployment page and re-run "验证 SSH."
"Voice or lip-sync output is wrong." Check the audio library: is the cloned voice actually assigned to the character? Is the line text present? IndexTTS 2.5 supports emotion control — an extreme emotion setting with a mismatched reference clip can produce odd results. Fall back to a neutral read, confirm it works, then dial the emotion back in.
"Batch generation partially fails." Run the failed cards individually. If they pass alone, the batch queue hit a transient issue — re-queue just those cards. If they fail alone too, treat each as an input problem per the steps above.
"Everything fails after a break." Check system logs first. Cloud instances get stopped, credits run out, tunnels drop. Reconnect, restart ComfyUI, then retry. This is the "have you tried turning it off and on" of AI video — and it works more often than anyone admits.
---
H2: Building a Debugging Habit That Scales
The creators who ship consistently aren't the ones who never see failures — they're the ones who read the logs before reacting. A few habits worth adopting:
- Keep AI logs and system logs open side by side during any batch run. You'll catch problems at card three instead of card thirty.
- Test one variable at a time. Change the prompt *or* the reference image *or* the cloud instance — never all three.
- Use single-shot generation as your unit test. It's the fastest way to prove whether a failure is local to one card or systemic.
- Write your own failure notes. A short text file per project, listing symptom → log evidence → fix, will save you hours on the next project.
- Know when to ask. If the logs point at something you can't resolve — a cloud mirror that won't appear, an SSH verification that keeps failing — reach out. The customer service WeChat is ymcandai, and the team also runs a Taobao shop (search 杨墨川AI) and posts tutorials on Douyin, Xiaohongshu, and WeChat Channels under 杨墨川. Support email is info@ymcdirector.com.
One last framing: logs are not error messages, they're evidence. An error message tells you something broke. A log tells you *where, when, and with what input*. Read it like a detective, not like a user clicking "retry."
Start with the system log to confirm the stage is standing, then read the AI log to see what the actors were told to do. Nine failures out of ten will resolve themselves into a clear, fixable cause — and the tenth is usually a cloud instance that needs a nudge. For setup details, deployment guides, and the latest workflow tips, the official site and blog are the places to bookmark.