Cineground AI workflows

AI workflows · Archive restoration

Restoring archival footage with AI: a production pipeline

How we restored 200+ archive shots for a feature documentary: the method, the rules that held up, and what went wrong.

By Arno Faure, Cineground, Montréal31 October 2026About 10 minutes to read
Download the PDFContact
200+archive shots restored
~46 hof human time, whole project
~14 minof human time per shot

LTX Restore Comfy Cloud

Plans
128 (63%)
Credits used
14,879
Total time
~9 h
Average per plan
~4 min

Topaz Starlight Precise 2.6

Plans
70 (35%)
Credits used
347
Total time
~10 h
Average per plan
~9 min

Seedance image to video

Plans
1
Credits used
700

Every shot goes through the same steps: export from the edit, preparation, a render, and a return into the timeline. For LTX and Seedance, a prompt is also written after looking at five frames; Topaz Starlight takes none. The sources were mixed (interlaced and progressive, three frame rates, tape and film), so preparing and checking the footage took a large share of the work, alongside the renders. The aim throughout was to automate as much as possible, so that nobody repeats the same task 198 times.

Key points

For reasons of confidentiality, the client is not named and the documentary’s own footage cannot be shown yet. To explain what we did, this note uses archive footage from the Prelinger Archives (Internet Archive) instead, except for the airport shot.

1Source material

The edit timeline runs at 23.976 fps. The archive did not. Sources came at 29.97, 25 and 30 fps, roughly half were interlaced, and field order differed by format. Handling these differences before any model touches the picture is what makes the later steps predictable.

Source typeTypical propertiesHandling
H.264 files from DV-style transfers29.97 fps, interlaced, bottom field firstDeinterlace, keep source frame rate
XDCAM MXFInterlaced, top field first, video start offset of 3 to 17 framesDeinterlace, measure the true first frame (section 3)
Tape transfers (VHS, U-matic)Noise, unstable bottom edge, non-square pixelsCrop the edge, correct pixel aspect, optional stabilisation
Web-sourced filesLow resolution, heavy compressionTreated as progressive, aspect ratio kept
Stills and photographsOne imageHeld as a clip of the shot’s exact length

2Pipeline overview

01ExportFrom the source file, ProRes 422
02AlignFirst frame found by picture correlation
03PrepareFrame count, crop, MP4 for LTX and Seedance

One engine per shot

LTX RestoreVery old archive, film. Five-frame review and prompt.
Topaz StarlightTape, 1980s. No prompt.
SeedanceMissing or damaged shots. Five-frame review, start and end images, prompt.
04Re-importPlaced automatically in the timeline (Resolve MCP + Claude)
05ValidateFirst and last frame, length, by a person
06PatinaFinal grade by our in-house colourist: degradations applied on the restored image
manifest.csv single source of truth: in/out points, frame counts, padding, status. Every automated step reads and writes it.

The unit of work is a shot, not a reel. Each shot is a row in manifest.csv with its source file, exact range, frame count, crop, padding and a status stored beside it. Shots are processed in batches of about 30, and a batch is validated by hand before the next starts.

3Export and frame-accurate alignment

3.1 Export

Each shot is exported from its source file, not from the timeline render, as Apple ProRes 422. The chain deinterlaces, corrects non-square pixels and resets timestamps. The shot keeps its native frame rate; DaVinci Resolve conforms it on re-import.

# representative command, parameters vary per shot
ffmpeg -i source.mxf \
  -vf "setpts=PTS-STARTPTS,yadif,scale=iw*sar:ih,setsar=1,crop=W:H:X:Y" \
  -c:v prores_ks -profile:v 2 -an shot.mov

setpts=PTS-STARTPTS matters: some H.264 files start their video stream one frame late, which appears as a black first frame in the edit.

ProRes 422 is a good master and a good input for Topaz. LTX and Seedance, however, only accept MP4, so the file is converted at the end, just before upload. The ProRes file stays the reference.

A step that turned out to be unnecessary. We exported every shot as ProRes because we expected to use Topaz on most of them, and for Topaz it was the right choice. In practice most shots went to LTX (128 plans against 70 for Topaz), and for LTX and Seedance the ProRes file was only an intermediate: the engine takes MP4, so the shot was converted straight afterwards. If we did it again, we would export MP4 directly for the shots we already know are going to LTX or Seedance.
ffmpeg -i shot.mov -c:v libx264 -crf 12 -pix_fmt yuv420p -an shot.mp4

3.2 Why the edit’s in-point cannot be trusted

Resolve counts fields, not frames, for interlaced sources, so source in and out points were doubled for 35 shots and had to be halved. XDCAM MXF files carry a video start offset that Resolve and ffmpeg count differently. In addition, the source range was often one frame short of the timeline length.

3.3 Measuring the first frame

For each shot we export three stills from the timeline (at frames 0, 3 and 6) and search the source for the matching position by picture correlation. Correlation is insensitive to brightness and contrast, which matters when the file has different levels. The exact length is derived from the timeline: n = ceil(timeline_frames × source_fps / 23.976).

Result on 175 manifest rowsCount
Aligned, high confidence162
To check by eye6
Low confidence2
Not verifiable (hidden under another track)5

A shot already exported is re-exported only if the new start is clearly better (correlation gain above 0.003).

The goal of all this measuring is automation: finding the right frame by script, for every shot, so that nobody has to scrub through a source by hand 200 times.

Failure case. For one shot, the stills taken from the timeline were stale, so the correlation reported a confident but wrong start. The editor spotted it on the picture (the action began later than it should). The fix was to use Resolve’s own source in-point for that file type. Lesson: a high score is not proof, so the first and last frame of a shot are checked by a person when needed.

3.4 Handles and black bars

Ten-frame handles were tried and dropped: on re-import the handle appeared only before the clip and the length came out at 18 frames instead of 20. Exports use exactly the used range. Black bars are detected per shot and cropped at export; they are never re-added.

4Preparation for LTX Restore

We used the LTX 2.5 IC-LoRA Restore model [1]. It runs on Comfy Cloud, the hosted version of ComfyUI, with the credits and compute time shown above. Its constraints drive the preparation:

One prompt, the same full text for every tile Tile 1 Tile 2 Tile 3 Tile 4 overlap, fused 960 × 544
Schematic. The output frame is cut into four tiles of 960 × 544 that overlap (hatched areas). Each tile is denoised separately with the same full prompt, and the overlaps are fused at every step. A word that is in the prompt can therefore be painted into a tile that does not contain it.

For tape sources sent to LTX, an optional variant crops the unstable bottom line, which costs about 1.4% zoom that must be compensated in the edit. An optional stabilisation pass (vidstab) helps on handheld shots and struggles when large areas move, such as waves or crowds.

tail_pad = (8n + 1) - n_frames          # frames cloned at the end
render  -> output has n_frames + tail_pad
cut the last tail_pad frames after the render

Workflow file

The ComfyUI workflow we ran on Comfy Cloud is available to download: ltx-2.5-restore-workflow.json (180 KB). It loads the LTX 2.5 22B distilled model and samples in 8 steps with euler, through a tiled sampler that fuses overlapping tiles after every denoising step. The prompt fields hold the Coney Island prompts from section 5: replace them with your own. Drag the file onto a ComfyUI canvas to open it.

5Prompting

5.1 Five frames first

This step is for LTX and Seedance only. Topaz Starlight takes no prompt, so it has no five-frame review and no prompt writing. Before a prompt is written, five frames spread across the shot are laid out as a strip, with a colour check (black and white, faded or colour). The prompt answers three questions: what is in the shot, what must not change, and where the model is likely to invent.

Five frames spread across one shot, laid out as a strip
The five-frame strip for one shot: first and last frame, and three in between. Example shot: a boardwalk at Coney Island, 1940. Click to enlarge.

5.2 Rules that held up

5.3 Positive prompt, an example

The prompt written for the shot in the strip above, 42 words (the rule is 20 to 40, so slightly long), following the rules: period, place, light, clothing, then the fixed tail. It names no colour and no camera move. The one line of text that must stay readable is quoted.

1940 documentary footage of a crowded Coney Island boardwalk in daylight, two soldiers in wool uniforms and peaked caps, men and women in period suits and dresses, a sign reading "OCEAN BATHS", natural faces, sharp photographic detail, natural grain, high resolution footage

5.4 Negative prompt

blocky compression, banding, mosquito noise, ringing, blurry, smeared,
ghosting, low resolution, plastic, oversharpened halos, warped faces,
logos, signage, lettering, modern vehicles, modern buildings, graffiti,
modern clothing

This base list is edited per shot:

Remove from the negativeWhen
logos, signage, letteringThe shot contains lettering that must survive
graffitiThe shot contains painted art
modern buildingsTowers or buildings of the 1980s appear
blurry, smearedBlur is wanted, for example an out-of-focus background
modern vehicles, modern clothingThe shot contains vehicles or clothing the model would read as modern, for example a spacecraft or spacesuits

5.5 Text

It depends on the size of the lettering: big letters are usually fine, small ones are usually gibberish. On a shot with a banner, the first prompt produced invented words for the small text. Three options worked: quote the exact text in the prompt for the one line that must be readable; keep the rest blurred through the prompt and negative; or leave the lettering unreadable and add it back in the grade. A prompt written without looking at the frames also described an action that was not in the shot, which is why the five-frame step exists.

6Choosing the engine

EngineBehaviourUsed for
LTX RestoreReconstructs much more detailVery old archive and footage shot on film
Topaz StarlightFaithful, restrainedFootage shot on tape, 1980s, where the picture must not be rewritten
Seedance (image to video)Generates motion from still imagesShots that are too damaged: a recreated start image and end image, then one described camera move

Topaz Starlight in practice

On footage shot on tape in the 1980s, where the detail is already there, Starlight recovers a surprising amount of information. The upscaled result can be a little too crisp, which is why the last step of the pipeline (below, in section 7) puts the patina back.

Example: an archive shot at 480p, and the result of Topaz Starlight Precise 2.6 (the model we used, in Topaz Video, 3× upscale, progressive) at 1080p. The picture is already well defined, so Starlight only has to add resolution and detail, and it keeps the look of the original.

Original
0:00 / 0:00
Original (480p)
Original (480p)
Starlight Precise 2.6
Starlight Precise 2.6

Topaz Starlight Precise 2.6, 3× upscale. The picture was already well defined: more detail and resolution, same colour and grain. Frame at 14 s.

Freeze frames: the monorail

Two moments of the shot where the difference is easiest to see. Click an image to enlarge.

Original, 14.0 s.
Original, 14.0 s. The monorail passing: soft, with low contrast.
Starlight Precise 2.6, 14.0 s.
Starlight Precise 2.6, 14.0 s. The same frame: sharper, with the same colour and grain.
Original, close-up.
Original, close-up. The lettering on the side of the monorail, at 14.8 s.
Starlight Precise 2.6, close-up.
Starlight Precise 2.6, close-up. “WALT DISNEY WORLD MONORAIL SYSTEM” is crisp and the panel seams are clear. The text is not invented: it was already in the image.

Share of plans and credits

EnginePlansShareCredits used
LTX Restore12863%14,879
Topaz Starlight7035%347
Seedance10.5%700
Other31.5%n/a

With Seedance, the prompt must describe only what really happens in the shot, and nothing else. The order of work is: first the start and end images, then the animation between them.

Case: LTX on a 1940 film shot

A boardwalk at Coney Island, 1940, 640 × 480, 185 frames (already a valid 8n+1 count, so no padding). The prompt is the one in section 5.3.

Original
0:00 / 0:00
Original
Original
LTX Restore
LTX Restore

Colourised without any colour in the prompt. Small lettering on the neighbouring buildings is rewritten. Frame at 2.5 s.

Limit. Most of the shot holds up well: faces, uniforms, the crowd. The main sign, quoted in the prompt, stays correct. Smaller lettering on the neighbouring buildings does not. It is mostly acceptable on the outward pass of the pan and gets worse on the return, where some signs turn into invented, garbled words (for example “OCEEANTHS” for “OCEAN BATHS”). Lettering that is not named in the prompt is rewritten, not preserved.

Freeze frames: the signs in the background

The top of the frame at six moments of the shot, original above and LTX below. Click an image to enlarge.

Freeze frame at 0.0 s, original above and LTX below
0.0 s. The big sign is rebuilt correctly. On the neighbouring building, the small sign already drifts: “OEAN BATRS”, “OCEEN BATH SUTHS”.
Freeze frame at 0.4 s, original above and LTX below
0.4 s. Same signs a few frames later. The soldier’s cap hides part of the big sign and it still reads “OCEAN BATHS”.
Freeze frame at 0.8 s, original above and LTX below
0.8 s. The pan speeds up: motion blur, nothing is legible in either version.
Freeze frame at 5.2 s, original above and LTX below
5.2 s. The return pass starts, still blurred. LTX keeps the blur and the signs are not readable yet.
Freeze frame at 5.6 s, original above and LTX below
5.6 s. The signs sharpen. The big one holds, but the neighbouring sign now reads “OCEEANTHS” and “ORNX STAT”.
Freeze frame at 6.1 s, original above and LTX below
6.1 s. Last frame. Same invented words on the neighbouring building: lettering that is not in the prompt is rewritten, not preserved.

Second case: a food stand, and what LTX gets wrong

Another shot from the same film: customers at a hot dog stand, then a close-up of sausages on a griddle. 640 × 480, 265 frames (8 × 33 + 1, so again no padding). Positive and negative prompts:

1940 footage of a hot dog stand at Coney Island, customers in shirtsleeves eating hot dogs, a vendor in a white apron, rows of sausages on a steel griddle, natural faces, sharp photographic detail, natural grain, high resolution footage
blocky compression, banding, mosquito noise, ringing, blurry, smeared, ghosting, low resolution, plastic, oversharpened halos, warped faces, logos, signage, lettering, modern vehicles, modern buildings, graffiti, modern clothing
Original
0:00 / 0:00
Original
Original
LTX Restore
LTX Restore

Colourised, with faces, clothing and food holding up well. Frame at 3 s.

Limits. Faces, clothing and the food hold up well, and the shot is colourised again although the prompt names no colour. The limit is camera shake: LTX does not stabilise heavily shaky shots, so they should be stabilised first for a better result.

Case: when Topaz and LTX both fail, recreate the shot with Seedance 2.5

An airport shot from 1986, soft and low in contrast. Three attempts, in order.

Original
0:00 / 0:00
Original
Original
Topaz Starlight
Topaz Starlight
LTX Restore
LTX Restore
Seedance 2.5
Seedance 2.5

The same shot at 0.8 s, four ways. Only Seedance writes “swissair” correctly.

Zoom on the lettering

The same fuselage, cropped from each version at about 0.8 s. Click an image to enlarge.

Fuselage lettering, Original
Original. Illegible: the lettering is a red smear.
Fuselage lettering, Topaz Starlight
Topaz Starlight. The letters sharpen into a wrong word, close to “SWISSGE”.
Fuselage lettering, LTX Restore
LTX Restore. Clean shapes, but one letter too many: “swiisrair”.
Fuselage lettering, Seedance 2.5
Seedance 2.5. Reads “swissair”, correct and crisp.

How the lettering was fixed

Neither Topaz nor LTX could be made to write “swissair”. With Seedance, the lettering is fixed before the animation starts: the recreated start and end images, shown just below, already carry it correctly (“swissair” on the fuselage, “KLM” and “Royal Dutch Airlines” on the other aircraft). They become the first and last frame of the clip, and the prompt ends with one instruction, “Keep the airline lettering unchanged and legible.” The clip then inherits the text from its two images instead of inventing it.

Start image for Seedance
4 Start image. Recreated from the first frame. Click to enlarge.
End image for Seedance
5 End image. Recreated from the last frame. Click to enlarge.

The prompt describes one camera move and states that the aircraft do not move. It also asks to keep the airline lettering unchanged.

@image1 is the first frame and @image2 is the last frame.

SHOT 1: Slow pan right
Wide documentary shot at an airport apron in the 1980s. The camera pans slowly to the right along a long grey concrete wall at real-time speed (24fps), with a very subtle handheld drift. The parked blue and white KLM Boeing 747 is progressively revealed from its nose to its fuselage while the white Swissair 747 with its red tail and white cross slides behind it on the left. Two thin poles in the foreground drift from right to left across the frame. The aircraft do not move. Overcast grey daylight. Keep the airline lettering unchanged and legible.

SFX: low airport ambience, distant jet whine. No Music.

7Re-import into the edit

Restored shots are placed automatically (DaVinci Resolve driven by Claude through an MCP server) on a dedicated track at their original record position, scaled to fit, and trimmed to the frame count in the manifest. Three details cost us time:

Last step: getting the patina back

The work has two steps. Step one is restoration: the AI recovers detail that the original did not show. Step two is getting the patina back. After the shots are back in the edit, our in-house colourist grades them and applies degradations: softness, grain, texture. The restored image is deliberately a little too clean, and that is the point. Once the detail has been recovered, it can be manipulated afterwards, in the grade, instead of being locked into the result of a single render. The shot ends up more readable and still looks like the footage it came from.

8Tracking and quality control

Each shot passes through five states: exported, rendered, imported, validated, issue. They are stored per shot and shown in a dashboard with a stage bar per batch. Assistants and technicians mark import and validation; shots with an issue get a red tag in the file system and a note. Shots whose alignment needs a human check get an orange tag until a person clears them.

Validation is manual and per batch: the editor checks the first and last frame and the length of each shot in the timeline. A batch is not closed until its flagged shots are cleared.

9Failure log

SymptomCauseFix
Black first frameVideo stream starts one frame late in some H.264 filessetpts=PTS-STARTPTS at the start of the chain
Wrong in-pointFields counted as frames; MXF start offsetHalve for interlaced sources; correlate stills
Confident but wrong startStale stills from the timelineRefresh the stills; check first and last frame by eye
Handles lost on re-importHandle only before the clip, length 18 instead of 20Export exactly the used range
Lost frame in the timelineExclusive end frame in the APIPass the full frame count
Invented letteringLTX cannot copy small or blurry textBlur it, quote it in the prompt, or add it back in the grade
Prompt describes an action not in the shotPrompt written without looking at framesFive-frame review before every LTX or Seedance prompt
Wrong frame count for the modelModel requires 8n+1Pad the tail, cut after the render

10Limits and what we would change

References

  1. Prelinger Archives, Internet Archive. Source of the archive footage used in the examples, except the airport shot.
  2. Comfy Cloud, hosted ComfyUI, used to run LTX.
  3. Lightricks, LTX-2.5-22b-IC-LoRA-Restore, model card, Hugging Face. Prompt guidance and frame constraints.
  4. Topaz Labs, Topaz Video (Starlight models).
  5. ByteDance Seed, Seedance 2.0.
  6. FFmpeg documentation: yadif, tpad, prores_ks.
  7. Blackmagic Design, DaVinci Resolve (scripting API documentation ships with the application).
  8. Model Context Protocol.

Credits

Written by Arno Faure. The team on the project: