Hi everyone,
I’ve been working on an ESP32-based monochrome composite video synthesizer called BLACK DOG SYNTH, and after a fairly absurd amount of PAL/NTSC testing I’ve hit a point where I’d really appreciate input from people who understand composite encoding better than I do.
I’ve also been following LoFi Future / //SimpleViz for a long time. The current SimpleViz hardware I’ve seen in recent posts clearly uses an ESP32, and the output looks remarkably clean on projectors and CRTs. I remember LoFi Future mentioning in Instagram stories that the SimpleViz code was intended to be open-sourced eventually, so if they happen to see this, information about that implementation would be incredibly useful too.
But this is not specifically a request for SimpleViz source — I’m also interested in understanding the underlying signal problem.
Current BLACK DOG architecture
MCU: classic dual-core ESP32 DevKit
Output: internal DAC / composite video
Logical framebuffer:
432×192, 8-bit grayscale
The final architecture currently has two actual low-level transports:
PAL transport
NTSC transport
and each one can run with burst either enabled or suppressed:
PAL MONO — PAL timing, burst suppressed
PAL COMPAT — identical PAL transport, burst enabled
NTSC MONO — NTSC timing, burst suppressed
NTSC COMPAT — identical NTSC transport, burst enabled
MONO and COMPAT for each standard are otherwise the same signal engine:
- same H/V timing
- same framebuffer
- same active window
- same luma levels
- same renderer
- same DMA/presentation model
The only intended difference is the physical colour burst waveform.
The strange part
MONO looks extremely clean on a CRT.
Dense 1-bit black/white graphics can be displayed without coloured halos.
But once valid colour burst is enabled, high-contrast monochrome edges can produce rainbow / RGB halos.
This is especially obvious in PAL.
NTSC COMPAT does it too, but much less strongly — at this point it is almost negligible in normal content.
PAL COMPAT can be quite obvious.
The interesting part is that this is strongly dependent on luminance contrast:
- black/white thin geometry produces false colour
- grey geometry in the SAME texture often looks completely clean
- a 16-level grayscale test looks clean at its boundaries
- disabling burst removes the RGB artefacts even on dense black/white pixel structures
The artefacts remain visible even when the renderer/animation is paused, so this is not caused by moving artwork.
Receiver testing
I’m testing on several very different receivers:
- analogue CRT
- Unisheen video recorder/display
- cheap AV→HDMI scaler based around MacroSilicon MS1858E
- very cheap Chinese projector
I also tested the projector with a real Sony PAL camcorder to verify that it genuinely accepts PAL.
Current COMPAT modes lock well on the digital receivers.
PAL became reliably accepted only after fixing the PAL subcarrier/burst relationship properly.
Major PAL breakthrough
Earlier PAL implementations had valid-looking H/V timing but the projector refused them.
What finally changed this was implementing:
- coherent PAL subcarrier phase rather than arbitrary per-line reset
- correct line-to-line PAL phase behaviour
- proper SCH relationship
- PAL field sequencing / 8-field behaviour
- appropriate vertical interval burst blanking
After that, the same projector which previously said NO SIGNAL immediately recognised PAL and displayed a stable image.
So I believe the remaining colour problem is different from the earlier PAL-lock problem.
Things already tested
This has turned into a rather large test campaign.
I have tried roughly a hundred-plus signal/profile variations across the project, including:
Timing / transport
- several independent PAL implementations
- several independent NTSC implementations
- M5GFX / Panel_CVBS-derived experiments
- Aquaticus-inspired monochrome cores
- Bitluni-derived tests
- PAL line/sample-count matrices
- horizontal frequency sweeps
- exact-H vs slightly offset H
- multiple subcarrier clock approaches
- width sweep approximately 320–480 logical pixels
- several framebuffer heights
- field-coherent vertical mapping experiments
Signal levels
- sync depth
- blanking level
- black level
- white excursion
- active-edge guards
These did not remove the artefact.
Burst
- fixed burst
- alternating PAL burst
- proper PAL/SCH sequence
- burst amplitude sweep
- sparse burst every 2/4/8 lines
- short burst clusters
Sparse burst progressively destroys digital receiver lock.
Lowering burst amplitude gave a particularly interesting result:
the receiver continued to lock down to surprisingly low burst amplitude,
but the false-colour artefacts became larger and uglier, rather than
disappearing.
At very low burst level the image finally loses lock.
My guess is that chroma AGC / poorer chroma reference is amplifying the
false chroma generated from luma.
Luma processing
I tested:
- broad 3-tap low-pass
- sharper 5-tap FIR
- sharper 7-tap FIR
- adaptive isolated-pixel suppression
- narrow IIR/Nyquist notch experiments
All of these provided useful evidence, but I rejected them as product solutions.
Strong enough filtering does remove the rainbow artefacts.
Unfortunately it also causes one or more of:
- obvious blur
- lost 1-pixel detail
- broken circles/lines
- ringing/ghosting
- large FPS loss
That is why I’m reluctant to “solve” this with framebuffer post-processing.
Important frequency observation
One earlier PAL transport used approximately:
17.75 MHz sample clock
with:
864 active output samples
and:
432 logical horizontal pixels
so each logical pixel was effectively represented by two output samples.
This places a 1-pixel alternating black/white pattern extremely close to
the PAL chroma subcarrier region.
That explains why pathological 1-pixel textures produce strong colour.
However, changing logical width did not eliminate the problem globally.
13.5 MHz experiment
I also tested a BT.601-style 13.5 MHz transport.
PAL still had coloured edge artefacts.
NTSC at 13.5 MHz produced something especially interesting:
perfectly regular vertical green/blue/red bands moving horizontally
across the image.
The colour bands were extremely periodic, and at geometry boundaries they
formed the same sort of coloured halos seen in the normal content.
The ordinary NTSC transport was dramatically cleaner.
This looked like a useful clue involving the sample grid/subcarrier
relationship rather than random renderer noise.
Raw output purity check
To make sure the renderer/palette was not secretly inserting a periodic
luma pattern, I built direct active-video probes.
For example, in PAL:
normal framebuffer flat field:
MIN = MAX = 64
direct active flat field:
MIN = MAX = 59
unique active sample values:
1
No 2–8 sample repeating modulation was found.
NTSC direct flat tests also produced one constant active level.
So a uniform grayscale framebuffer really does become constant DAC samples.
This makes me think the artefact is not a hidden renderer/palette bug.
Current interpretation
The evidence currently points toward classic cross-colour / Y-C ambiguity:
- MONO without burst is visually clean.
- Enable valid burst and the receiver enables its chroma decoding.
- High-frequency B/W luminance near the chroma band can then be decoded
as colour. - Lower-contrast grey detail contains less energy and is often visually clean.
- Strong luma filtering removes the colour but destroys the image.
This seems physically reasonable.
What I’m trying to understand is how small commercial/D.I.Y. digital
monochrome generators such as //SimpleViz apparently avoid the ugly version
of this while still producing a signal that projectors happily accept.
Questions
I’d really appreciate thoughts on any of these:
1. For a deliberately monochrome composite generator, is emitting a normal
PAL/NTSC colour burst actually the wrong approach if high-frequency 1-bit
detail is important?
Is there a standard/commonly used way to remain widely compatible with
modern decoders while indicating monochrome?
2. Is there a way to generate a burst/reference that keeps digital receiver
lock without causing the chroma decoder to interpret luma as colour?
My sparse/amplitude experiments suggest probably not in the naive way.
3. How do real monochrome composite generators typically control
cross-colour without visibly softening pixel graphics?
Is the answer usually:
- deliberate luma bandwidth limiting
- analogue reconstruction/filtering
- coherent sample clock design
- different effective horizontal resolution
- comb-friendly line structure
- something else?
4. How important is the relationship between DAC/sample clock and Fsc when
synthesising CVBS directly from a microcontroller?
The 13.5 MHz NTSC moving RGB bands made this look much more important than
I originally assumed.
5. Is there a particularly sensible sampling scheme for direct ESP32 DAC
generation?
For example, is staying at an integer multiple of Fsc substantially better
than using a convenient independent sample clock?
6. Could the ESP32 internal DAC / zero-order hold itself be contributing a
significant amount of spectral energy around chroma frequencies?
Would a modest analogue reconstruction network after the DAC make a major
difference even if the digital samples are mathematically clean?
7. For anyone familiar with the current ESP32 //SimpleViz:
I’d be incredibly interested to know:
- Arduino-ESP32 vs native ESP-IDF / custom low-level code
- actual framebuffer resolution
- video sample clock
- I2S/DAC vs GPIO/resistor output
- whether burst is present in its B/W output
- progressive/288p-style vs true interlace
- analogue output network/filtering
I remember LoFi Future mentioning in Instagram stories that the SimpleViz
firmware/code was intended to be open-sourced eventually. If that code or
even an older ESP32 revision is available anywhere, I’d love to study it.
I’m not asking anyone to debug the whole BLACK DOG codebase. At this point
I mainly want to understand whether I’m missing a fundamental composite
encoder design technique before I declare the current PAL/NTSC engines done.
If it would help, I can post schematics, actual sample/timing values, photos
of the artefacts, or the full test matrix.
Thanks — this forum has been one of the few places where discussions go
deep enough into actual analogue video signal behaviour that I figured this
was worth asking here.










