-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathDEMO_VIDEO_SCRIPT.txt
More file actions
173 lines (154 loc) · 8.35 KB
/
Copy pathDEMO_VIDEO_SCRIPT.txt
File metadata and controls
173 lines (154 loc) · 8.35 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
====================================================================
NEARBY MOOD MAP -- AMD AI DevMaster Hackathon Demo Video Script
Track 1: Development of Multimodal Content Creation Tools
Target length: 3:30-4:30 (fits the 3-5 min recommended window)
====================================================================
BEFORE YOU RECORD -- REQUIRED CHECKLIST
--------------------------------------------------------------------
[ ] Recording on the actual AMD Radeon GPU / ROCm machine, not a dev
box. The rules require core inference to run locally on AMD
Radeon GPU -- this is graded (20 pts) and checked in the video.
There is no non-ROCm fallback or testing override in the code --
PixArt simply refuses to run without a real ROCm PyTorch runtime.
[ ] Confirm ROCm is actually active before recording:
curl http://127.0.0.1:8000/api/health
Look for "pixart": {"ready": true, "rocm": true, ...}. If "rocm"
is false, do not record yet -- fix the environment first.
[ ] GEOAPIFY_API_KEY and TICKETMASTER_API_KEY are set in .env (live
providers -- without these the map falls back to fewer results).
[ ] Backend running: uvicorn api:app --host 127.0.0.1 --port 8000
[ ] Browser open to http://127.0.0.1:8000 with the window sized so
the 3D map and side panels are all visible without scrolling.
[ ] Pick TWO real search examples ahead of time so the "diversity"
section (step 6 below) isn't improvised on camera -- e.g. a
"hungry" search in a city with restaurants, and a "tired" or
"want live music" search in a city with cafes/venues. Confirm
both actually return good results BEFORE recording.
[ ] Have a terminal window ready to alt-tab into for step 2 and
step 5 (showing the backend actually running / ROCm generation
happening is part of the grading criteria: "from command line
/GUI to final result").
====================================================================
SCRIPT
====================================================================
--------------------------------------------------------------------
[0:00-0:20] HOOK / WHAT THIS IS
--------------------------------------------------------------------
ON SCREEN: Start on the empty map page (before any search), showing
the search panel.
SAY:
"This is Nearby Mood Map -- a local discovery app that turns a
one-line description of your mood into a personalized 3D map of
nearby places, with every place illustrated by a unique AI-generated
image created live on an AMD Radeon GPU using PixArt-Sigma, running
on the ROCm software stack."
--------------------------------------------------------------------
[0:20-0:45] THE CONCEPT / WHY GENERATED IMAGES
--------------------------------------------------------------------
ON SCREEN: Stay on the search panel; point at the mood textarea.
SAY:
"Instead of showing generic stock photos or map pins, every place on
this map gets its own representative hero image, generated on the
fly by a diffusion model -- so the map itself becomes a piece of
generative art tailored to how you're feeling and what's actually
nearby."
--------------------------------------------------------------------
[0:45-1:15] INPUT -- MOOD + LOCATION
--------------------------------------------------------------------
ON SCREEN:
- Type a real mood into the textarea, e.g.:
"I'm hungry and want something comforting nearby"
- Select a destination from the dropdown (or click "Use my
location" if recording somewhere with real nearby venues).
- Show the search radius field.
- Click "Find nearby places".
SAY (while doing the above):
"I'll describe how I'm feeling, pick a location, and set a search
radius. Under the hood this hits our FastAPI backend, which queries
Geoapify Places and Ticketmaster for real nearby venues and events
within that exact radius."
--------------------------------------------------------------------
[1:15-2:15] PROCESSING -- SHOW THE ROCm GENERATION HAPPENING
--------------------------------------------------------------------
ON SCREEN:
- Let the status line update live: "Searching nearby places...",
then "Generating hero image 3 of 8 on ROCm...".
- Briefly alt-tab to the terminal running uvicorn and show the
PixArt inference log lines scrolling (diffusion step progress),
then tab back to the browser before it finishes.
SAY:
"This is a background job, not a single blocking request -- the
backend starts a job immediately and the browser polls it for live
progress. For every nearby place it returns, PixArt-Sigma -- an
open-source, ~0.6-billion-parameter text-to-image diffusion model --
generates a clearly-labeled representative image, with inference
running locally on the AMD Radeon GPU through ROCm. Nothing here is
a stock photo or a cached placeholder -- if ROCm inference isn't
available, the API fails loudly instead of faking a result."
--------------------------------------------------------------------
[2:15-3:15] OUTPUT -- THE 3D MAP
--------------------------------------------------------------------
ON SCREEN:
- Show the finished 3D map: "you are here" pin, POI billboards
with their freshly-generated images, the ground-plane map image,
and the recommendation card.
- Read the recommendation text out loud from the card.
- Click "Yes, show directions" OR click a POI's billboard directly
to open its detail panel (name, distance, category, rating if
available, image label).
- Open the collapsible POI navigator panel on the right and press
W/S or the arrow keys to browse a couple of places, showing the
camera panning to each one.
SAY:
"Here's the result: a live 3D map centered on my location, with
every nearby place rendered as its own AI-generated billboard. The
top recommendation is surfaced automatically with a one-line reason,
and I can browse every other place with this navigator panel --
keyboard or click -- each one panning the camera and showing that
place's own generated image and details, with a direct link to
OpenStreetMap directions."
--------------------------------------------------------------------
[3:15-3:50] EFFICIENCY TOUCH -- CLIENT-SIDE RADIUS NARROWING
--------------------------------------------------------------------
ON SCREEN: Drag the search radius input down to a smaller value and
show the ring and POIs update instantly.
SAY:
"Narrowing the search radius doesn't trigger a new backend call or
new GPU generation at all -- every place and image within the
original radius is already loaded, so this just re-crops the scene
client-side, instantly."
--------------------------------------------------------------------
[3:50-4:20] DIVERSITY OF OUTPUT -- SECOND EXAMPLE
--------------------------------------------------------------------
ON SCREEN: Run the pre-tested second search (different mood and/or
location from the checklist above) and let it finish generating.
SAY:
"To show this isn't one lucky result -- here's a completely different
mood and location: [read the new mood text]. New real venues, a new
generated image for every one of them, all produced fresh on the same
ROCm pipeline."
--------------------------------------------------------------------
[4:20-4:45] WRAP-UP / TECH RECAP
--------------------------------------------------------------------
ON SCREEN: Return to the finished map, wide shot.
SAY:
"Nearby Mood Map: real nearby places from live provider APIs, a
unique PixArt-Sigma image generated per place on an AMD Radeon GPU
through ROCm, and an interactive 3D web map to explore them --
built entirely on an open-source model stack, with core inference
running locally, no closed-source generation API involved. Thanks
for watching."
====================================================================
NOTES FOR EDITING
====================================================================
- If PixArt generation for a full search takes a while on camera, it
is fine to speed up / cut the waiting portion of section 3 in
post -- just keep a few real seconds of visible progress-log text
and the terminal cutaway so the grader can see it's really running,
not faked.
- Keep the terminal cutaway text legible (zoom in) -- graders are
explicitly asked to check "actual execution performance on an AMD
Radeon GPU, from command line/GUI to final result."
- Do not include any generated image that looks like it depicts a
real, identifiable person -- the prompts already guard against
this, but do a quick visual check on whatever examples you record.