HM3 — generated logo concepts
Why these exist. The first two artefacts could only produce marks a typeface can set. These were invented from a written brief instead, so the field includes forms no font contains. That is the whole value here, and it is the only value here.
Every one needs redrawing. These are raster images from a diffusion model. Curves are approximate, symmetry is approximate, and lettering is unreliable. Treat a card as an argument for a shape, then rebuild that shape in vector before it goes anywhere near production.
Tags. keep means the idea earns a vector redraw, maybe means it needs a second pass on the brief, and reject means the concept failed on its own terms. Lockups also carry a spelling tag, because a mark that does not read HM3 is not a candidate at all.
ASymbol concepts
Ten marks with no lettering at all. A symbol is what a diffusion model can actually contribute, because shape has no spelling to get wrong. The question for each of these is whether the idea underneath is worth redrawing by hand, since the pixels themselves are disposable.
3.1Signal burst reject
Three concentric arcs radiating from a corner point: a transmission crossing distance.
Drawn cleanly, and it would scale beautifully, which is exactly the problem. This is the RSS glyph and it is also the wifi glyph, two of the most recognised icons on any screen, so HM3 would spend the next decade borrowing a shape rather than owning one. Nobody wins that argument.
3.2Iris M maybe
A camera iris whose blades leave a letter M standing in the negative space.
The M does resolve, and tying the initial to the equipment is a genuine idea. But the blade gaps are uneven and read as cracks, so the disc looks shattered rather than mechanical. Worth a redraw with true rotational symmetry before it is judged.
3.3Waveform three reject
The numeral three built from three stacked signal waveforms.
Failed on its own terms. Three waveforms were stacked, but no numeral emerged from them, so the count is only legible if someone counts. It also reads as audio rather than television, and the ragged wave edges will not hold below about 40px.
3.4Tower maybe
A transmission tower stripped to three tapering bars with a signal dot above.
Clean, confident, and it survives to very small sizes. The tapering stack reads as a pyramid or a tree at least as readily as a tower though, and the floating dot does more work than the tower does. Narrowing the taper would decide it.
3.5Layers reject
Three squares overlapping in perspective, the overlaps lighter, for three layers of media.
Elegant and completely unusable. The whole idea depends on transparency, so with the accent switched off the overlaps vanish and the mark becomes a grey blob. It also lands squarely in generic enterprise-software territory.
3.6Negative M3 keep
A solid block with the M standing and the numeral three cut out of its counter.
The strongest thing generation produced. One solid shape, so it holds in one ink, knocks out of footage and reduces to a favicon without a redraw of the idea, and it carries both the M and the numeral in a single form, which no option in either vector round manages. The counters want opening and the three wants regularising. Redraw this one.
3.7Shutter three maybe
A rotary shutter whose curved blades resolve into the numeral three.
The numeral does read, and a rotating blade is a real television reference rather than a borrowed one. Against it: three tonal values and an outer keyline mean it needs four colours to work, and the curl of the blade gives it a late-nineties feel that a tighter geometric redraw would have to fix.
3.8Gauge reject
A semicircular level meter with graduated ticks and a needle.
Handsome and wrong. This is the dashboard icon shipped by every analytics product on the market, and it needs three tonal steps plus hairline ticks, so it fails one ink and dies well above favicon size. It also argues for measurement rather than for television.
3.9Tape loop reject
A ribbon of videotape twisted once into a closed figure.
The most beautiful drawing in the batch and the least useful, because it reads unmistakably as a figure eight or an infinity sign. Wrong number. The form also depends on tonal shading to show the twist, so flattening it to one ink destroys the only thing that made it interesting.
3.10Grid three maybe
The numeral three built from a grid of squares, like a broadcast character generator.
Reads instantly as a three and the titler reference is genuinely theirs rather than borrowed. The problem is scale: the gaps between cells close up by about 32px and the mark turns into a solid blob, so it would need a coarser grid with fewer, larger cells.
BLockup concepts
Eight full lockups where the model was asked to set HM3 itself. Letterforms are the part it gets wrong, so every card below records whether the mark actually spells HM3 before it says anything about the design.
3.11Hex emblem maybe spells correct
A compact hexagonal emblem holding an H, with HM3 set beside it in a heavy grotesque.
Spelling is right and the lockup proportion is sound. The emblem itself is the problem: a letter inside an angular hexagon is the default esports and crypto badge, and the split light-dark treatment adds a bevel the palette does not want.
3.12Stacked crest reject spells correct
A symbol above, HM3 centred beneath it.
Spelling is right and the stack is well proportioned. The crest reads as a gaming clan badge, all points and armour, which is the wrong century and the wrong industry for a shop selling on live television and Amazon.
3.131986 geometric keep spells correct
HM3 in a wide, confident geometric sans of the mid-eighties television era.
The best pure wordmark generation produced. Wide, even, quietly period-correct without dressing up as a costume, and it holds together as one horizontal block. One flaw to fix in the redraw: the 3 has a flat squared top terminal that reads momentarily as a B, which is fatal in a three-character mark.
3.14Stencil slate maybe spells correct
HM3 in a stencil broadcast-slate face, held between two thin rules.
The rules do real work, framing the mark the way a slate frames a take, and the stencil breaks give it a utility feel that suits a production company. But the breaks land arbitrarily rather than on a system, the three characters sit at three different tonal values for no reason, and the hairline rules will not survive a favicon.
3.15Monoline circle reject spells unclear
HM3 interlocked in one continuous monoline stroke, inside a thin circle.
The characters share strokes so aggressively that a stray horizontal runs along the baseline, and the mark reads as HML3 or HMB before it resolves into anything. A three-character logo that needs a second look has already lost. The hairline weight disappears below about 40px anyway.
3.16Knockout chip keep spells correct
H and M solid, the 3 reversed out of a platinum square sitting tight against the M.
Worth noting that this arrived from a written brief with no reference image at all, and it converges almost exactly on the Tally construction from the first vector round. Two independent routes landing on the same answer is the strongest signal in this entire exploration. The chip is slightly wider than square and the letterforms are approximate. The idea is confirmed.
3.17Mirror rule maybe spells correct
Ultra-condensed HM3 with a single rule slicing through the characters.
The model misread the brief and reflected the whole wordmark below the rule instead of slicing through it. The accident is more interesting than the instruction: a mark and its own reflection across a horizontal line reads as a broadcast signal and its return. It needs the reflection resolved into something deliberate rather than a mirrored duplicate.
3.18Serif with star maybe spells correct
HM3 in a high-contrast modern serif with a small geometric symbol above.
The serif setting is genuinely elegant and makes the heritage argument better than any sans here. The symbol undercuts it: a four-point sparkle is the current shorthand for artificial intelligence and appears on half the products shipped this year. Drop the star, keep the letterforms, and this becomes the Didot direction from the typeface axis with better spacing.