Video wird geladen...
Video konnte nicht geladen werden
I gave OPUS 5.5 Access to: - Uni 1.1 for image generation - Seedance 2.5 - Davinci Resolve Studio for Grade and editing I also gave it a Suno track I had already made. It made this entire video 🤯 More below 🧵
42,231 Aufrufe • vor 1 Tag •via X (Twitter)
45 Kommentare

OPUS first took inventory of my very loose idea and analyzed the audio track I had given it. It then built itself a beat sheet that had three acts:

I wired up an API call to Uni 1.1 from @LumaLabsAI so OPUS could generate and iterate on the look by itself. Here are the 5 key visuals it came up with to use with Seedance 2.5

Next it generated a single Seedance 2.5 video using those 5 images + this prompt below (attached video is raw output of Seedance 720p): [Generation Goal] Generate one continuous, heavily cut cinematic action sequence. During a naval battle in the drizzle, a lone mechanical gunner walks out onto the forward deck of a warship. His massive coil cannon powers up, and at the climax he fires one colossal beam into an enemy mothership of impossible scale that hangs over the sea. The build is micro-detail cutaways intercut with battle B-roll, the huge hit comes at around 17 seconds, and the firing payoff lasts about ten seconds before settling into darkness. [Reference Material Roles] @ Image 1 through @ Image 5 are key visuals, not keyframes. They define the look of the subjects, props, effects and world. None is a first or last frame, and none should be reproduced as a static hold. @ Image 1 defines the gunner's clawed metal boot, the black cloth leg wraps, the frayed hem of the olive-grey oilskin cape and the wet steel deck plating. @ Image 2 defines the gunner's full design: the hooded olive-grey waxed oilskin, slate-navy gunmetal armour, the long heavy half-cape and the deep braced crouch. It also defines the cannon (near-black squared spine, dark coil collars with orange-red glowing gaps, blunt squared muzzle block, dense receiver cluster beside the head, thick looping cables), the forward deck with its railings and gantries, and the flat overcast drizzle lighting. @ Image 3 defines only the beam and the sea-boil effect: a blinding white beam with a thin orange-red filament fringe, and the sea beneath it flash-boiling into towering walls of white steam. Do not use its framing as a fixed shot. @ Image 4 defines the inside of the hood: the angular gunmetal faceplate in deep shadow and the single thin horizontal orange-red slit of light. @ Image 5 defines the enemy mothership: a single dark slab-sided hull tens of kilometres across that fills the sky and runs past both edges of the frame. Its underside is a city of plating, antenna forests and turret blisters with thousands of dim red lights, pressing into an unbroken low cloud ceiling above a fleet reduced to specks. [Subjects and Relationships] There is exactly one gunner and exactly one cannon throughout. The gunner is entirely mechanical, with no skin, no face and no mouth, and he never speaks. The cannon is three times his height and only ever rests on his right shoulder. The beam only ever leaves the squared muzzle. The mothership is the only enemy and always hangs over the sea ahead of the bow. Friendly ships are slate-grey industrial warships. The cloud ceiling stays unbroken throughout, with no hole, ring or parting. [Visual Style] Photoreal live-action science fiction with a practical costume and real heavy machinery. Soft, shadowless overcast daylight in fine drizzle. The palette is cool and desaturated: slate navy, gunmetal, olive-grey and pale sky white, with the orange-red slit and coil gaps as the only warm notes. Subtle film grain and a muted grade. The beam is the only thing that ever blows out the frame; when it fires, it throws hard cold-white light across every wet surface. Serious, grounded, monumental cinematography. [Beat Map and Cinematography] 0–3 seconds: Cut in hard on an extreme wide shot from low on a neighbouring ship's deck, with a wet railing soft in the corner. A warship in heavy grey swell fires a full broadside, and orange muzzle flashes punch through the drizzle. Cut to a 24mm angle looking straight up: tracer lines climb into the cloud, and a flash briefly reveals the mothership's dark underside resting on the whole ceiling. <Naval guns boom in a rolling broadside> <rain hisses on steel> <tracers crack upward> <a deep sub-bass groan from the cloud> 3–6 seconds: A hard cut to a macro shot inches off the deck: the clawed boot slams onto a wet plate, throwing up a crown of water, and the frayed cape hem swings into frame. Cut to a mid-distance destroyer taking a hit: a fireball and a tall column of spray. <A single heavy metal footstep> <water splashes> <a distant explosion thuds across the water> 6–10 seconds: A low shot from knee height behind the gunner as he walks toward the bow, the soaked cape dragging. Deck crew in foul-weather gear press against the railings and turn to watch him pass. Cut to a long-lens shot from sea level: the mothership sinks slowly out of the cloud, cloud pouring off its edges like waterfalls, the fleet below reduced to specks. Cut to an 85mm extreme close-up in profile inside the hood as the orange-red slit flickers on. <Heavy footsteps on steel> <wind snaps the wet cape> <muffled crew shouts> <a vast low drone from the mothership> <a small electronic ignition tick> 10–12 seconds: A tight shot as the cannon's heavy dark cradle slams onto his right shoulder pad, knocking water off the armour, with the rest of the weapon out of frame. Cut to a 200mm long-lens shot from far back: the gunner is a dark silhouette at the forward rail, the cannon's long mass reaching past the bow, the mothership looming beyond. <A crushing metal impact> <the deck shudders> <cables thud against steel> 12–15 seconds: A power-up montage of shallow-focus macro cuts that never shows the whole cannon. A coil gap bleeds orange-red, and raindrops flash into steam on contact. A battered pressure gauge needle pins in the red. The cables jolt as the mechanical fingers clamp tighter on the grip. The coils ignite one after another toward the muzzle, a line of orange light racing away from camera. <A rising electrical whine> <rain sizzles into steam> <capacitors thunk as they charge> <escalating electrical crackle> 15–17 seconds: The held breath. A close shot along the glowing cannon: the drizzle stops mid-air, and thousands of raindrops hang motionless around the barrel. Near silence. An off-screen male voice over the ship's radio shouts in Japanese: {撃て!} Cut to a symmetrical head-on shot into the squared muzzle, the hood and slit barely visible behind the receiver, as the slot aperture fills with blinding white. <Sound drops to a thin high ringing> <a single radio click> <the whine peaks into a sharp charging inhale> 17–20 seconds: The payoff. At the instant of discharge, cut to the hero wide three-quarter side view from slightly below, matching @ Image 2. A colossal straight white beam, far wider than the cannon itself, erupts from the muzzle and runs out of frame. The deck flashes white, recoil drives the gunner back, the cape rips backward, and a shockwave ring blasts the rain flat. Then cut on the beat to the same instant three more times: over his right shoulder looking down the barrel as the beam tears toward the mothership; low at deck level as the boots skid backward, gouging the steel while the plates buckle; and an extreme close-up of the hood as the slit flares. <A colossal discharge thunderclap> <a tearing energy roar> <metal shrieks as the boots skid> <deck plates buckle> <a violent pressure wave of spray> 20–24 seconds: An extreme wide shot from a distant warship: the massive beam crosses the frame from a tiny ship up into the mothership's underside, flashing the seascape cold white. Cut to a low shot skimming the sea beside the beam, matching @ Image 3: the sea flash-boils into two towering walls of steam, and a small warship heels in the violence. Cut to an extreme wide shot from sea level: the beam punches deep into the mothership, and a chain of white explosions races across kilometres of hull while debris the size of ships falls toward the sea. <A continuous roaring beam> <the sea boils with a violent hiss> <a titanic impact rumble> <cascading detonations> <huge metal structures tear> 24–27 seconds: A wide shot as a shockwave rolls back across the sea toward camera, pressing the waves flat, and slams into the deck, throwing rain and cape sideways. Cut to the gunner, still braced and still firing, a black silhouette against solid white beam light. <A deep shockwave boom> <gale-force wind and spray> <the beam's roar holds> 27–30 seconds: The beam cuts out. The coil gaps fade to dark, and thick steam pours off the barrel into the returning drizzle. Far beyond the muzzle, the mothership burns and sheds debris into the sea. Cut to an extreme close-up inside the hood, matching @ Image 4, as the orange-red slit slowly dims to nothing and the frame settles into darkness. <The beam cuts out> <cooling metal ticks> <steam hisses> <rain patters back> <a low power-down> <silence> [Audio] No music. The only audio is the specified diegetic sound effects and the single off-screen Japanese line {撃て!}. The gunner makes no vocal sounds. There is no narration, and no subtitles or on-screen text appear. [Maintain Consistency] Keep exactly one gunner, one cannon and one mothership. Keep the hood, single slit, armour, cape, boots, and the cannon's squared spine, coil collars, muzzle and receiver identical across every shot. The cannon always rests on his right shoulder and aims ahead of the bow, and the beam comes only from the muzzle. The overcast drizzle world returns whenever the beam is not lighting the frame, and the cloud ceiling stays unbroken.

I then manually upscaled 720 to 4k and gave the video back to it and told it to edit and post process and color grade in Davinci which is the OP video of this thread. Here is the exact instructions I gave it for the davinci section: ok I went with the other video (I prompted it manually with the reference images) Here's what I want you to do next. I have davinci resolve studio open on my mac right now and I have the video + the audio already preloaded for you. You task is to edit the video so that it is a perfect cinematic trailer. There may be parts we have to trim, re-order, audio management etc. One thing I want the "FIRE" part to be clear. It's around approximately 16s or 16.5s right before the gun fires. You need to also deal with the music track to make it fit properly etc. I am leaving this up to your creative decisions! Make it fucking epic as hell. Finally, I want you to color grade and post process the entire thing (filmic grain, effects..... everything a real professional studio would do here). I am going to paste in a summary below but by no means do you need to adhere 1:1: DaVinci Resolve Studio — Seedance 2.5 Full-Video Colour Grading Brief Objective Perform a professional, cinematic colour grade on the entire supplied Seedance 2.5 video in DaVinci Resolve Studio. The goal is to achieve a cohesive, intentional, filmic image while retaining natural detail, smooth tonal transitions, believable skin tones, controlled highlights, and rich but realistic colours. Prioritize image quality and consistency over applying a heavy stylized look. Source assessment Before grading: Inspect the entire source video, not just the opening frame. Identify the source resolution, frame rate, codec, bit depth, chroma subsampling, and available colour metadata. The source is expected to be H.264, 8-bit, 4:2:0, at 24 fps. Determine whether the source is SDR or HDR and identify its colour space and transfer function where possible. Inspect representative frames for clipped highlights, crushed blacks, banding, compression artifacts, unusual skin tones, inconsistent white balance, and AI-generated visual anomalies. Preserve the original frame rate and duration unless explicitly instructed otherwise. Do not assume that the footage is log-encoded or that its colour metadata is correct simply because it is AI-generated. 2. Colour management Use DaVinci Resolve Studio's high-precision processing pipeline. Preferred starting configuration for an SDR master: Colour science: DaVinci YRGB Color Managed. Timeline working colour space: DaVinci Wide Gamut / DaVinci Intermediate. Output colour space: Rec.709 Gamma 2.4 for a conventional SDR master. Input colour space: identify the source's actual encoding. If metadata is absent or unreliable, inspect the image and verify the interpretation before committing to a grade. Tone mapping and gamut mapping: use appropriate output transforms to produce a controlled Rec.709 image. If the source is confirmed to be conventional Rec.709 SDR, assign the appropriate Rec.709 input colour space. Do not apply a log-to-Rec.709 LUT to footage that is already display-referred Rec.709. If the source's colour characteristics cannot be established confidently, document the assumption and compare the result against the original. 3. Data levels In the Media Pool, inspect Clip Attributes → Data Levels. Start with Auto. Verify that the image has plausible black and white levels. Use waveform monitoring to check for unexpected clipping or lifted blacks. Only override Auto if there is evidence that the source is being interpreted incorrectly. Do not select Full as a quality enhancement. The goal is correct interpretation, not an arbitrary conversion between full and video levels. 4. Bit depth and intermediate files The source is 8-bit H.264 4:2:0. Resolve's internal processing provides substantially greater precision than the source encoding. Do not describe an 8-bit-to-16-bit conversion as recovering missing colour information. Do not transcode the source unnecessarily before grading. If an intermediate is needed for VFX, compositing, or an external application, use a suitable high-bit-depth format such as ProRes 4444 or an appropriate high-bit-depth image sequence. Choose the format according to the actual workflow and alpha-channel requirements. A higher-bit-depth intermediate can preserve subsequent processing more accurately, but it cannot restore source detail that was never present. 5. Grading approach Grade the complete video for a consistent cinematic appearance. Primary correction Establish balanced exposure and contrast. Correct obvious white-balance or colour casts. Set convincing black levels without crushing shadow detail. Preserve highlight detail wherever it exists in the source. Maintain natural midtones and believable skin tones. Avoid excessive saturation, especially in reds and oranges. Creative look Develop a restrained, cohesive cinematic look appropriate to the footage. Use deliberate contrast and tonal separation. Keep highlights controlled and shadows rich without making them muddy. Introduce subtle colour separation only where it improves the image. Avoid excessive teal-and-orange treatment, artificial skin tones, or exaggerated saturation unless the scene specifically calls for it. Preserve the intended mood and visual continuity of the original. Shot-to-shot consistency Review every shot and transition. Match exposure, white balance, contrast, and saturation across adjacent shots. Use shot-specific adjustments where necessary rather than forcing one identical correction onto every shot. Watch for flickering colour, inconsistent exposure, and changes in skin tone. Avoid abrupt visual changes at cuts. Protect the source Because the source is 8-bit 4:2:0: Avoid unnecessarily aggressive shadow recovery. Avoid extreme saturation pushes and large secondary corrections that reveal compression or chroma artifacts. Watch gradients, skies, dark backgrounds, and smooth areas for banding. Use noise reduction only when it demonstrably improves the image; avoid smearing fine details. Do not sharpen aggressively or invent texture to conceal compression. Do not use AI enhancement or generative modifications that alter the content unless explicitly authorized. Use scopes as well as visual judgment. Review the image at normal playback size and inspect problem areas at 100% magnification. 7. Export Create a high-quality master suitable for further editing and archival. Preferred master: Resolution: 3840 × 2160 if the approved project is intended to be 4K. Frame rate: preserve the source frame rate of 24 fps. Colour space: Rec.709 Gamma 2.4. Codec: ProRes 422 HQ or an appropriate DNxHR high-quality intermediate. Bit depth: use the highest practical precision supported by the selected codec. Audio: preserve the original audio, synchronization, and duration unless instructed otherwise. Data levels: choose the correct setting for the selected codec and delivery requirements. Embed appropriate colour-space and gamma metadata where supported. Do not upscale a 1280 × 720 source to 4K merely to label it 4K. If the supplied source is 720p, establish whether an approved 4K upscale exists before selecting the delivery resolution. If the source is already 3840 × 2160, preserve that resolution. Also create a delivery-ready H.264 or H.265 version if requested, using an appropriate bitrate and bit depth. 8. Quality control Before delivery: Watch the entire graded video from beginning to end. Check every cut for colour or exposure discontinuities. Confirm there are no accidental black frames, dropped frames, altered timing, or missing audio. Check highlights, shadows, skin tones, gradients, and saturated colours. Compare the finished grade with the source. Confirm that the rendered file has the intended resolution, frame rate, codec, colour-space tags, and duration. Verify that the rendered file looks consistent with the Resolve viewer. Deliverables Provide: The final graded master. A delivery-ready compressed version, if requested. The DaVinci Resolve project file, with media links managed appropriately. A brief report identifying the colour-management settings, export settings, any assumptions made about the source, and any source-quality limitations. Final instruction Use professional colourist judgment throughout. The objective is the best technically sound and aesthetically cohesive grade the source can support—not an artificial attempt to make 8-bit footage behave like original 16-bit or RAW footage. If a technical uncertainty materially affects the image, test the alternatives on representative frames before applying the decision to the entire video. Do not silently make destructive changes to the source.

AI video agent SaaS maybe over!

Insane!

I need to learn how you all are connecting Opus to these tools.

Wow!! Incredible…opus is crazy awesome

Indeed very cool

Ha damn. That's quite epic. How much manual iteration, tweaking, regens are you doing? Or was this essentially a one shot with manual steps in between?

Essentially a one shot. It did generate 20ish “mood” images and I told it to use those 5. That was the extent of my input

Ha kk. One step closer to the end game then. :P

Looks really good, well done!

Thank you. This is my first time giving OPUS the keys to DaVinci Studio and I have to say I’m really impressed with it. I 100% recommend trying it

What…

Giving it Resolve is the part most people leave out, and grading is where a sequence either becomes one piece or stays a folder of clips. Did it hold a look across the whole thing or shot by shot?

It sliced the video into shots and built a fusion graph for each individual shot

So how much credits did you spend?

It was 1 single Seedance gen so like $10 or whatever 720, 30s is It generated about 25 images total at 4 cents an image approximately The suno track was free as I used the v6 mini to make it Claude was part of monthly sub So like..$15? (seedance expensive)

How much quota usage it burn for this whole process 👀

TBH next to nothing. Also UNI 1.1 is about 4 cents per image so you can just blast

So, claude Opus 5.5 usage is very low even in this whole workflow. The reason I'm asking because a small 3 minute analyzing with Opus take upto 3% to 5% of quota usage and here I suppose the agent did analyze what it generated for the overall context.

Yeah but it packs contact sheets at like 2fps

What do you mean by contact sheets at 2fps. You mean embed with the content?

No direction whatsoever? Love the visuals, I haven’t tried that image model yet.

There was direction in my opening prompt. I just used voice-2-text and just brain slopped out some idea.

insane work, bravo

I’ll pass along the kind words to homie OPUS 😂

lol, this whole process was hands off? that's wild. opi cookin.

This looks amazing! Did you edit the track manually to fit the video or was it also done by Opus?

Opus did that too. I just generated a 30s track in Suno v6

Epic

😱

Opus got the keys to Seedance, Resolve, and a Suno track, then somehow walked out with a finished cut.

Budget?

I'm interested in Uni 1.1 the generations your getting out of it look great!

I did the same - just one difference, I ONLY used the song as prompt, nothing else - and this came out: 👾

Well that looked fucking awesome

This needs to be Fallout 5.

I love these videos. Hope they were free

Thank you so much for sharing such detailed information. 🙏🙂↕️ Only if you don't mind, can you please let me know what kind of computer configuration you have for generating such amazing output?

You are a meat proxy.

Siempre me preguntaban por trabajos que iban a sobrevivir a la IA y yo decía “editor de vídeo”. Hay que ver lo equivocado que estaba.

Dick

That's ridiculously cool 😯

