MEMOACTS 2026 ONLINE INTENSIVE 21.08

MODULE 03

Ethics and Techniques of Intervening in Archival Footage

Thirty seconds of one archival film, put through four operations of rising intervention. Every clip on this page was produced on a local ComfyUI install; each carries what model made it, at what settings, on what machine, in what time, and what was done to the source.

Where this sits

5 min

Creator's workflow

1Idea and the topic
2Research to collect and verify materials
3Decide how to present the document — you are here
4Write a screenplay and record the voiceover
5Assemble the visuals
6Polish: subtitles, visual effects, sound design
7Pre-publishing screening and evaluation

Schedule

14.08	01 Introduction and Finding the Focus
18.08	02 Research Phase — AI-powered or AI-distorted
21.08	03 Ethics and Techniques of Intervening in Archival Footage + Creating a Moodboard
25.08	04 Editing as a Way of Thinking Through a Fragmented Reality
27.08	05 Tools Implementation: Choose What Works or Build Your Own
02.09	06 Production: Visual Effects and Sound Design
07.09	07 Final Screening and Wrap Up

The reel — Signed After Midnight

8 min

Format: 1080×1920, 2:49, 20 shots
Narration: recorded human voice, passed through unre-encoded
Subtitles: verbatim from the script; timings computed by forced alignment (stable-ts), text never transcribed from audio
Archive in it: at 1:01, KAPFILM footage, Museum Berlin-Karlshorst, in-point 0:40 — full width with bands, at native size, no enlargement
Built with: memoacts_core + ffmpeg, from shots.csv

Visual identity and the moodboard

22 min · designer

Presented live. Slides supplied separately.

The film

10 min

Not available online

The source film is shown in the session only. Everything below describes it; the reel and the four ladder steps are on this page.

The source, whole — 7:36. Tempelhof, the signing, the banquet. The working cut starts at 04:14

Source: Museum Berlin-Karlshorst, KAPFILM, 7:36
Working cut: 04:14–04:44 — the signing. 901 frames at 30 fps
Frame: transfer is 1280×800; picture is 1068×800 at x=106. Pillarboxed; the cut is de-pillarboxed by crop=1068:800:106:0
Audio: −91 dB mean and max end to end, every sample in one histogram bin. Dropped from the working cut
Colour: the scan carries a constant warm offset, a few levels more red than green throughout
Also present: the museum's own burned-in trilingual captions elsewhere in the reel, absent from this stretch

Discussion

The crop, the transfer, the tint and the museum's captions all arrived as decisions someone else made. Does that make your changes more defensible, or less?
This stretch was chosen partly because the museum's captions are absent from it. Is choosing the convenient thirty seconds already an edit?
Nothing in this film is signed — no cameraman, no editor, no scanning technician. What can you honestly say about provenance?

Four levels

5 min

Level	Operation	Model	Cost
1 — quotes	text composited over the frame	none	CPU
2 — sound	audio generated from text, mixed to picture	Stable Audio Open 1.0	6 s / layer
3 — upscale	1068×800 → 2136×1600	4x_foolhardy_Remacri (ESRGAN)	3.66 s / frame
4 — colour	instruction edit, then chroma transplanted onto the original	Qwen-Image-Edit (Apache-2.0)	117–156 s / frame

All four ran on a local ComfyUI install, RTX 3090 Ti, 24 GB.

1 — quotes

15 min

Three quotes over the working cut

Model: none
Render: ffmpeg, libass filter, one pass over 30 s. Seconds, on CPU
Graph version: VHS_LoadVideo → DrawText+ → VHS_VideoCombine, one run per quote, ComfyUI local
Text source: transcribed by hand from the signed English copy of the Act of Military Surrender. Not read off the footage
Done to the archive: nothing. Pixels unmodified; text composited on top
Caption band: measured 14–65 of 255 across the cut, so white text with a shadow stays legible without a plate

Discussion

A caption is a verbatim quote from a real document. Does putting it over a different moment of the film make it a lie? Where does timing stop being editing and start being a claim?
The reel carries the museum's own burned-in captions. If yours contradicts theirs on the same frame, whose survives?
Does a quote belong on screen if you cannot also show where it came from?

2 — sound

15 min

The working cut with four generated layers mixed under it

Model: Stable Audio Open 1.0, open weights, with T5-base as a separate text encoder
Where: ComfyUI local, RTX 3090 Ti
Sampler: 50 steps, cfg 4.98, dpmpp_3m_sde_gpu, exponential, denoise 1.0
Generation time: ≈6 s of GPU per layer, independent of requested length
Done to the archive: picture untouched. The source audio track was discarded; all four layers are generated from text
Mix: ffmpeg — atrim, adelay, per-layer gain, amix normalize=0, loudnorm I=-20 TP=-2 LRA=11

The four layers

Layer	Sec	Seed	Prompt
bed	32	550919	Ambience of a crowded hall: many men murmuring quietly at a distance, low indistinct voices, occasional cough and chair creak, reverberant, no music.
paper	14	220805	A single sheet of stiff dry paper lifted from a wooden table, held, turned over, and set down again. Close, dry, detailed, in a quiet room.
cameras	16	880431	Sharp mechanical camera shutter clicks and clacks, film advance ratchet winding, a flashbulb pop, spaced out with silence between them, recorded in a large reverberant room. Dry transient clicks, no continuous tone.
steps	12	440108	A heavy wooden chair scraping back on a parquet floor, then several slow leather-soled footsteps on parquet in a large echoing room. Close and dry.

Negative prompt on all four: music, melody, singing, instruments, rhythm, beat, speech, dialogue, narration. Bed and cameras are second takes — the prompt was rewritten, the seed changed with it.

Discussion

Nobody captions a room tone. Should they? What would a sound credit look like on a thirty-second video, and would anyone read it?
The narration in the reel is a recorded human voice. If it were synthetic, would you disclose that? Does the answer change for the crowd murmur?
Stable Audio Open is trained largely on Freesound's CC0 library — sounds people recorded and gave away. Whose coughs are in that 1945 hall?

3 — upscale

15 min

Same 1000×800 patch from each — left: lanczos enlargement · right: Remacri

One frame cut down the middle — left: lanczos · right: Remacri

Model: 4x_foolhardy_Remacri, ESRGAN architecture
Where: ComfyUI local, RTX 3090 Ti, --disable-pinned-memory --cache-none
Graph: VHS_LoadVideo → UpscaleModelLoader → ImageUpscaleWithModel → ImageScaleBy → SaveImage, 8 frames per submission
Generation time: 3.66 s per frame · 901 frames · ≈55 min total
Done to the archive: the model runs at its trained 4× (4272×3200), then one lanczos reduction to 2× (2136×1600). Edges are reconstructed; film grain is removed as noise
Baseline: the same frames enlarged with lanczos only, which adds no detail

Detail added, measured on one face at the delivered 2×

Model	Detail	Over baseline
lanczos (baseline)	2.27	—
4x_foolhardy_Remacri	2.34	+3 %
4xlsdirplus_v1	2.35	+4 %
4x-UltraSharp	2.51	+11 %
4x_NMKD-Siax_200k	3.19	+41 %

Faces in the source are about twenty pixels tall. Remacri is not available on Comfy Cloud; 4x-UltraSharp and 4x_NMKD-Siax are.

Discussion

The grain is part of the recording. Is removing it restoration, or retouching?
Your platform offers two upscalers and both add grain to skin that has none. You cannot install the careful one. Do you use what is there, or not upscale at all?
Legibility against invention: at native size and speed nobody watches. Is that worth it? The reel above answered no.

4 — colour

15 min

The same seconds of the signing, three times.

1 — original

2 — colour laid over the original

3 — each frame re-invented by the model

Model: Qwen-Image-Edit fp8 (Apache-2.0), with qwen_2.5_vl_7b_fp8_scaled and qwen_image_vae
Where: ComfyUI local, RTX 3090 Ti, one frame per submission
Sampler: 20 steps, cfg 2.5, euler, simple, denoise 1.0, ModelSamplingAuraFlow shift 3.0, at 1068×800
Generation time: 117–156 s per frame, warm
Playback: 10 fps, 6 s. All three clips are the same span of film
Clip 1: the source frames, unmodified
Clip 3: the model's output as returned. Denoise 1.0 is required — at 0.6, 0.4 and 0.25 the frame survives but almost no colour arrives
Clip 2: the source frame's L channel kept, a and b taken from the generation and fitted to the source luminance with a guided filter (radius 12, eps 1e-3). Shapes, edges and grain come from the film; only chroma comes from the model
Colourisation prompt: names olive-brown tunics, gold shoulder boards, red collar tabs, field grey, dark panelling, tungsten light — a historical claim typed by the operator

Frame-to-frame variation, against the film's own noise floor

Region	Variation
tunic	4×
face	4.6×
blank plaster wall	17.6×

Nothing in the pipeline enforces consistency between frames; each frame is generated independently. There is no colourisation model on Comfy Cloud — clip 2 is produced by tools/module03_colorize.py, outside ComfyUI.

Discussion

The model returned a plausible modern photograph where a document had been, and reported success. Nothing failed. What in your process would have caught it? Name the step.
Keeping the film's luminance means no shape can move. Does that make the colour honest, or only make the invention harder to see?
The blank wall varies 17.6× the film's own noise. Is that a defect to fix, or the only visible sign the viewer gets that this is generated?

Close

5 min

Level	What was added to the record
1	Words from a document. No pixel of the image changed.
2	A soundtrack. The source carried none.
3	Texture and edges no camera recorded; the grain removed.
4	Colour — and, in clip 3, the image itself.

One sentence each: which level is the last one you would use on your own subject, and what would you have to write down for the next person to trust it?

Homework

Both installed before 25.08

1 — Start a reference board

are.na — make an account and open a channel for your own film. Put references in it as you find them: images, frames, typography, colour, layouts, anything that answers the question the moodboard block asked, how should this look. Bring the link to module 04 on 25.08.

2 — Install a video editor

shotcut.org — free, open source, Windows / macOS / Linux.

3 — Install an audio editor

audacityteam.org — free, open source, Windows / macOS / Linux.

Both installed before 25.08.