AI video credits have a way of disappearing. You start the month fine, and by the second week you’re watching good ideas die because every experiment cost full price. I’ll be real — I burned through that cycle myself more than once before I changed how I generate.
The fix wasn’t a cheaper tool or a better prompt. It was splitting every shot into two passes: a cheap draft pass to prove the shot works, and an expensive finish pass that only approved shots ever reach. It’s how film productions have always worked — rehearsals before the expensive camera rolls — and it maps almost perfectly onto AI generation.
The Two-Pass Idea
Pass One asks a single question: does this shot work? Does the action read, does the timing land, does the composition make sense, does the character do the thing the story needs? Pass Two asks a different question entirely: how beautiful can this proven shot be?
Most creators run one pass that tries to answer both questions at once, at maximum cost, several times in a row. That’s the expensive habit. When a full-quality generation fails on story — the character walks the wrong way, the joke doesn’t read, the camera move fights the action — you’ve paid flagship prices to learn a rehearsal lesson.
Two-pass generation splits the risk. Drafts are cheap enough to be disposable, so you explore freely. Finals are expensive, so you spend them only on shots that already earned it. In a proper production pipeline this isn’t a trick — it’s just a stage gate. Nothing moves to finishing until it’s approved in draft. I run my Seedance 2.5 and Veo work this way now, and I would not go back to single-pass even if credits were free. Because the decisions get better, not just the bill.
Pass One: Motion And Story At Draft Quality
The draft pass runs at whatever your tool’s cheap setting is: draft mode if it has one, lower resolution if it doesn’t, fewer seconds if the shot allows. The prompt stays the real prompt. This matters — a draft made with a lazy prompt proves nothing about the real shot. Write the prompt you’d ship, generate it cheaply, and judge what comes back on story terms only.
What I’m looking for in Pass One:
- The action reads without explanation. If I have to tell someone what just happened, the shot failed, regardless of beauty.
- Timing. Does the beat land where the storyboard said it would? A storyboard-first workflow gives you the panel to compare against — draft on screen, panel next to it. Match or miss, it’s obvious.
- Camera behavior. Does the move help or fight? A slow push-in that arrives too late reads just as clearly at draft quality.
- Performance shape. The broad expression, the turn of the head, the pause before the line. Not facial detail — the shape of the performance.
- Continuity sanity. Right character, right wardrobe direction, right props in the right hands. The cheap stuff to check is the expensive stuff to fix later.
Generate wide in this pass. Two, three, four drafts of a tricky shot is normal and healthy at draft prices. I keep the drafts in a clearly named folder — project_shot_drafts — because draft files that wander into the finals folder cause exactly the confusion you’d expect. On multi-shot sequences from Seedance 2.5, I’ll often draft the entire sequence end to end before finishing a single shot. Story problems usually live in the joins between shots, and the joins only show up when you watch the whole thing.
What You Are Allowed To Judge In A Draft (And What You Are Not)
This section exists because drafting fails in a predictable way: creators judge drafts like finals and kill good shots for crimes the final pass would never commit.
You may judge: story clarity, motion quality, timing, composition, camera logic, gross continuity errors, whether the character’s broad performance works, whether the shot matches its storyboard panel. These are structural qualities. They’re resolution-independent. And they carry over to the final.
You may not judge: skin detail, eye sharpness, fine hair behavior, fabric texture, small background items, subtle facial micro-expression, color richness. Drafts smear all of it. I’ve watched creators regenerate a draft six times chasing skin quality that the draft engine was never going to show them. The take was approved on motion in generation one. They just didn’t know they’d approved it.
The discipline is a sentence I say out loud during draft review: “Am I judging the shot, or the pixels?” If it’s pixels, it waits for Pass Two. Writing this judgment down helps too — a one-line verdict per draft (“Shot 7 draft 3: motion approved, ignore face, finish this one”) is the core of handling client revisions without chaos, because clients also need to learn which pass they’re looking at. Show a client a draft labeled clearly as a motion draft, and they’ll comment on motion. Show them the same file without the label, and you’ll get notes about skin.
Pass Two: Finals Only For Approved Shots
Pass Two is boring by design. And boring is good. Take the approved draft’s exact prompt — same words, same references, same first frame if you used one — and run it at full quality: full resolution, the tool’s best mode, proper seconds. The approved draft is your map; Pass Two just re-renders the territory beautifully.
A few working habits for the finish pass:
- Change one thing at a time. If the final needs a tweak the draft didn’t have — slightly tighter framing, different light — change that and only that. Stacked changes make failures impossible to diagnose.
- Keep the draft beside the final. Play them side by side. The final should be the same shot, prettier. If it wandered into a different shot, your references weren’t locked tightly enough.
- Use image-to-video for hero shots. For close-ups that matter, finish from an approved still frame. The storyboard-first habit of approving the frame first makes this natural: the frame is already sitting in your project approved and waiting.
- Stop at good. The finish pass tempts endless re-rolls because each one is slightly different. Judge finals against the checklist, not against an imaginary perfect take. Two full-quality attempts per approved shot is my usual ceiling; a third means something structural is wrong and I’m back in draft territory with a clearer prompt.
Finished clips go into Premiere Pro, get their trims and grade, and the draft files get archived — not deleted. Drafts are evidence of decisions. When a client asks why the character turns left in shot nine, the draft sequence with its approval notes answers in seconds.
The Credit Math That Makes This Obvious
I won’t invent prices, because tool pricing moves around and your plan isn’t mine. But the ratio is stable enough to plan around, and you can plug in your own numbers in a minute.
Count your approval rate honestly. On story-driven sequences, roughly how many generations does it take before a shot is approved? For most creators I talk to, story approval takes two to four attempts per shot once you count wrong motion, wrong timing, and wrong framing. At single-pass full pricing, every one of those learning attempts costs flagship money. If a final-quality generation costs, say, five times a draft on your tool — and ratios in that neighborhood are common — then approving in draft and finishing once can cut the cost of a tricky shot to a fraction, even after paying for the final. Three drafts plus one final at that ratio costs less than two full-price guesses. And you made better decisions, because you saw more attempts.
The math gets kinder the more experimental the shot. A wild camera move or a physical comedy beat might take six drafts to find. Six full-price generations hunting for a gag is how budgets die. Six drafts plus one finish is a Tuesday.
Run the numbers on your own last project. Count total generations, count shipped shots. The gap between those numbers is what two-pass generation refunds you.
Tools Without Draft Modes: The Manual Version
Not every tool offers an honest draft mode, but the method doesn’t need one. The manual version uses whatever cost levers your tool exposes:
- Lower the generation resolution for Pass One. 720p drafts judge motion just fine — this is exactly the territory of generation versus delivery resolution.
- Shorten the draft. If the final shot is six seconds, a four-second draft often proves the motion. Extend or regenerate at full length for the finish.
- Draft a still, not a clip. The cheapest draft of all is a single frame. If the frame is wrong, the clip would have been wrong at fifty times the price. Approve frames, then animate.
- Turn off the premium toggles. Whatever your tool calls its quality enhancers, they belong to Pass Two.
What doesn’t work as a manual draft: generating at full quality “but only once, carefully.” That’s not a pass — that’s hope with a bigger bill. The whole point is iteration volume at low cost. If your tool truly has no cheaper setting — some don’t — then draft with stills and short clips and accept a leaner version of the method. It’s still cheaper than finishing unapproved shots.
What Changes When You Actually Stick To It
A month into running two-pass properly, three things shifted in my work that I didn’t predict. First, I started attempting harder shots. When a wild idea costs a draft to test instead of a flagship generation, you test it. The ceiling of what I try rose because the floor of each attempt got cheap — and some of my best-performing shots began life as “this will probably fail” drafts. Single-pass pricing had been quietly censoring my ideas, and I hadn’t noticed until the censorship lifted.
Second, my prompts got better, faster. A draft is a prompt diagnostic: when three drafts all fail the same way — the character enters from the wrong side, the prop appears late — the prompt is telling you exactly what’s ambiguous in it. I rewrite between passes now, which means Pass Two inherits a tested instruction instead of a first guess. Most creators iterate prompts by regenerating at full price and squinting. Drafting turns prompt-writing into a conversation with quick, cheap replies.
Third, deadlines calmed down. A client revision that used to mean “regenerate the expensive thing and pray” now means “draft the change, get approval, finish once.” Clients approve motion drafts more honestly than finals, too — show someone a polished render and they critique colors; show them a labeled draft and they critique the thing you’re actually deciding. The two-pass split improves communication, not just cost. And on paid work, the communication is worth more.
None of this requires discipline of the monk variety. It requires a drafts folder, honest labels, and one rule you don’t break under deadline pressure: nothing unapproved gets finished. Tape that sentence where you can see it during generation. It’s saved me more money than every other habit in this article combined.
Keep it simple enough that you’ll actually do it on deadline: draft folder, verdict notes, finish only what’s approved. Two-pass generation isn’t about being stingy. It’s about spending your credits on beauty after the story already works — which is the only order that ever made sense.