home / www.linum.ai / item 47141107
Model Comparison
100% sign agreement
Model Editorial Structural Class Conf SETL Theme @cf/meta/llama-3.3-70b-instruct-fp8-fast lite 0.00 ND Neutral 0.90 0.00 No human rights theme @cf/meta/llama-4-scout-17b-16e-instruct lite 0.00 ND Neutral 0.90 0.00 Technology Development deepseek/deepseek-v3.2-20251201 +0.35 +0.40 Moderate positive 0.08 -0.13 Scientific Sharing claude-haiku-4-5-20251001 +0.29 +0.25 Neutral 0.08 0.06 Open Science & Knowledge Access meta-llama/llama-3.3-70b-instruct:free ND ND — — — —
Section @cf/meta/llama-3.3-70b-instruct-fp8-fast lite @cf/meta/llama-4-scout-17b-16e-instruct lite deepseek/deepseek-v3.2-20251201 claude-haiku-4-5-20251001 meta-llama/llama-3.3-70b-instruct:free Preamble ND ND 0.14 0.18 ND Article 1 ND ND ND ND ND Article 2 ND ND ND ND ND Article 3 ND ND ND ND ND Article 4 ND ND ND ND ND Article 5 ND ND ND ND ND Article 6 ND ND ND ND ND Article 7 ND ND ND ND ND Article 8 ND ND ND ND ND Article 9 ND ND ND ND ND Article 10 ND ND ND ND ND Article 11 ND ND ND ND ND Article 12 ND ND ND ND ND Article 13 ND ND ND ND ND Article 14 ND ND ND ND ND Article 15 ND ND ND ND ND Article 16 ND ND ND ND ND Article 17 ND ND ND ND ND Article 18 ND ND ND ND ND Article 19 ND ND 0.63 0.52 ND Article 20 ND ND ND ND ND Article 21 ND ND ND ND ND Article 22 ND ND ND ND ND Article 23 ND ND ND ND ND Article 24 ND ND ND ND ND Article 25 ND ND ND ND ND Article 26 ND ND 0.33 0.18 ND Article 27 ND ND 0.83 0.66 ND Article 28 ND ND ND ND ND Article 29 ND ND ND ND ND Article 30 ND ND ND ND ND
Summary Open Science & Knowledge Access Advocates
This technical blog post documents Linum's development of an Image-Video VAE, emphasizing open-sourcing of model code, weights, and experimental logs. The content engages substantively with Articles 19 (free expression and information sharing), 26 (education and technical learning), and 27 (scientific participation), advocating for democratic access to AI research artifacts and transparent methodology in scientific progress.
Article Heatmap
Preamble: +0.18 — Preamble P Article 1: ND — Freedom, Equality, Brotherhood Article 1: No Data — Freedom, Equality, Brotherhood 1 Article 2: ND — Non-Discrimination Article 2: No Data — Non-Discrimination 2 Article 3: ND — Life, Liberty, Security Article 3: No Data — Life, Liberty, Security 3 Article 4: ND — No Slavery Article 4: No Data — No Slavery 4 Article 5: ND — No Torture Article 5: No Data — No Torture 5 Article 6: ND — Legal Personhood Article 6: No Data — Legal Personhood 6 Article 7: ND — Equality Before Law Article 7: No Data — Equality Before Law 7 Article 8: ND — Right to Remedy Article 8: No Data — Right to Remedy 8 Article 9: ND — No Arbitrary Detention Article 9: No Data — No Arbitrary Detention 9 Article 10: ND — Fair Hearing Article 10: No Data — Fair Hearing 10 Article 11: ND — Presumption of Innocence Article 11: No Data — Presumption of Innocence 11 Article 12: ND — Privacy Article 12: No Data — Privacy 12 Article 13: ND — Freedom of Movement Article 13: No Data — Freedom of Movement 13 Article 14: ND — Asylum Article 14: No Data — Asylum 14 Article 15: ND — Nationality Article 15: No Data — Nationality 15 Article 16: ND — Marriage & Family Article 16: No Data — Marriage & Family 16 Article 17: ND — Property Article 17: No Data — Property 17 Article 18: ND — Freedom of Thought Article 18: No Data — Freedom of Thought 18 Article 19: +0.52 — Freedom of Expression 19 Article 20: ND — Assembly & Association Article 20: No Data — Assembly & Association 20 Article 21: ND — Political Participation Article 21: No Data — Political Participation 21 Article 22: ND — Social Security Article 22: No Data — Social Security 22 Article 23: ND — Work & Equal Pay Article 23: No Data — Work & Equal Pay 23 Article 24: ND — Rest & Leisure Article 24: No Data — Rest & Leisure 24 Article 25: ND — Standard of Living Article 25: No Data — Standard of Living 25 Article 26: +0.18 — Education 26 Article 27: +0.66 — Cultural Participation 27 Article 28: ND — Social & International Order Article 28: No Data — Social & International Order 28 Article 29: ND — Duties to Community Article 29: No Data — Duties to Community 29 Article 30: ND — No Destruction of Rights Article 30: No Data — No Destruction of Rights 30
Aggregates
Editorial Mean +0.29 Structural Mean +0.25 Weighted Mean +0.39 Unweighted Mean +0.39 Max +0.66 Article 27 Min +0.18 Preamble Signal 4 No Data 27 Volatility 0.21 (Medium) Negative 0 Channels E: 0.6 S: 0.4 SETL ℹ +0.06 Editorial-dominant FW Ratio ℹ 60% 12 facts · 8 inferences
Theme Radar
Foundation Security Legal Privacy & Movement Personal Expression Economic & Social Cultural Order & Duties Foundation: 0.18 (1 articles) Security: 0.00 (0 articles) Legal: 0.00 (0 articles) Privacy & Movement: 0.00 (0 articles) Personal: 0.00 (0 articles) Expression: 0.52 (1 articles) Economic & Social: 0.00 (0 articles) Cultural: 0.42 (2 articles) Order & Duties: 0.00 (0 articles)
HN Discussion
6 top-level · 3 replies
Hi HN, I’m one of the two authors of the post and the Linum v2 text-to-video model (
https://news.ycombinator.com/item?id=46721488 ). We're releasing our Image-Video VAE (open weights) and a deep dive on how we built it. Happy to answer questions about the work!
This seems like a great model to experiment fine tuning with original art, given it’s relatively small and with open license. Is that a fair assessment?
Thanks for the great write up and making it available to us all.
Nice summary! I missed the mention of EQ-VAE when it comes to generation quality. Tiny trick, huge impact! Have you tried it?
Very nice well written article!
The kind that I like so much on HN. It tickle your mind but is still clear enough for an advanced beginner.
its cool to see the iterative improvements to your model laid out, but for everything that workedm i imagine there were at least a million other things you also tried but didnt work out. whats your process of trying these different techniques/architectures? do you just wait for one experiment to finish and visually inspect the results everytime. seems hard since these take a while to train. how do you shorten the feedback loop in this space?
This is very cool thanks for sharing
yep, Apache 2.0! so anyone's welcome to download and hack away
No questions but I appreciate the write-up! Thank you for sharing.
Hadn’t seen that before! Seems very in line with what with the broader points about regularization. In table 4 they show faster convergence in 200 epochs when used alongside REPA. I’d be curious to see if it ended up beating REPA by itself with full 800 epochs of training — or if something about this new latent space, leads to plateauing itself (learns faster but caps out on expressivity). We’ve seen that phenomena before in other situations (eg UNET learns faster than DiT because of convolutions, but stops learning beyond a certain point).
Editorial Channel
What the content says
+0.45
Medium Advocacy Framing
Content strongly advocates for participation in scientific advancement and cultural life through transparent research publication and open knowledge sharing.
FW Ratio: 60%
Observable Facts
Page explicitly releases 'Model code', 'Model weights', and 'experiment logs' to public without apparent access restrictions Content documents detailed technical research process, including experimental failures and methodological decisions Organization commits to continuing this practice: 'we're approaching our next VAE in 2026', suggesting sustained commitment to open science Inferences
Open-sourcing of research artifacts directly enables public participation in scientific advancement and cultural production Transparent publication of methodology and results contributes to the scientific commons and cultural heritage of knowledge +0.35
Medium Advocacy Framing
Content demonstrates commitment to free expression and information sharing by openly publishing technical research, methodology, code, and experiment logs.
FW Ratio: 60%
Observable Facts
Page explicitly states 'we're open-sourcing our Image-Video VAE, our experiment logs' Direct links provided to download 'Model code' and 'Model weights' without apparent access restrictions Technical details and research process are transparently documented, including failures ('fighting through months of NaNs') Inferences
Open-sourcing of code and experiments supports Article 19's protection of freedom to seek, receive, and impart information Transparent publication of research methodology and results enables public scrutiny and informed discourse about AI development +0.20
Medium Advocacy
Content advocates for open-sourcing AI research and democratizing access to technical knowledge and model artifacts.
FW Ratio: 60%
Observable Facts
Page states 'we're open-sourcing our Image-Video VAE, our experiment logs, and a key finding' Content includes links to 'Model code' and 'Model weights' for public access Article discusses technical research process transparently, sharing both successes and failures ('fighting through months of NaNs') Inferences
Open-sourcing of code and weights aligns with principles of democratic access to knowledge and scientific transparency Transparent discussion of research failures and challenges supports informed public understanding of AI development +0.15
Medium Advocacy
Content promotes access to technical education and scientific knowledge through open-sourcing research artifacts and detailed explanatory documentation.
FW Ratio: 60%
Observable Facts
Page provides freely available code and trained model weights, reducing cost barriers to technical education Content includes visual reconstructions with paired original/reconstruction labels, suggesting accessibility consideration Technical documentation describes research methodology and findings in detail, supporting educational engagement Inferences
Open-sourcing of trained models and code lowers barriers to technical education for learners without institutional affiliation Transparent publication of research process supports public understanding of AI development and methodology ND
Not engaged. Content does not address equal and inalienable rights of humans.
ND
Not engaged. Content does not address non-discrimination.
ND
Not engaged. Content does not address security of person.
ND
Not engaged. Content does not address slavery or servitude.
ND
Not engaged. Content does not address torture or cruel treatment.
ND
Not engaged. Content does not address right to recognition as person.
ND
Not engaged. Content does not address equal protection before the law.
ND
Not engaged. Content does not address effective remedies for rights violations.
ND
Not engaged. Content does not address arbitrary arrest or detention.
ND
Not engaged. Content does not address fair trial or due process.
ND
Not engaged. Content does not address criminal proceedings or presumption of innocence.
ND
Not engaged. Content does not address privacy, family, home, or correspondence.
ND
Not engaged. Content does not address freedom of movement or residence.
ND
Not engaged. Content does not address asylum or right to nationality.
ND
Not engaged. Content does not address right to nationality or its change.
ND
Not engaged. Content does not address marriage or family.
ND
Not engaged. Content does not address property rights.
ND
Not engaged. Content does not address freedom of thought, conscience, or religion.
ND
Not engaged. Content does not address freedom of assembly or association.
ND
Not engaged. Content does not address political participation or democratic governance.
ND
Not engaged. Content does not address social security or social services.
ND
Not engaged. Content does not address work, employment, or labor rights.
ND
Not engaged. Content does not address rest, leisure, or reasonable working hours.
ND
Not engaged. Content does not address health, food, clothing, housing, or medical care.
ND
Not engaged. Content does not address social and international order.
ND
Not engaged. Content does not address duties to community or limitations on rights.
ND
Not engaged. Content does not address prohibition of destruction of rights.
Structural Channel
What the site does
+0.40
Medium Advocacy Framing
Site provides unrestricted public access to model code and weights; open-source architecture removes barriers to accessing and disseminating technical information.
+0.35
Medium Advocacy Framing
Open-source release of code and weights enables broad participation in scientific progress and technical innovation; transparent methodology supports scientific culture.
+0.15
Medium Advocacy
The site provides downloadable model code and weights, enabling knowledge sharing and reducing gatekeeping of technical resources.
+0.10
Medium Advocacy
Basic accessibility features present (alt text for images); open-source model access removes economic barriers to technical learning.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not engaged. Content does not address equal protection before the law.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
ND
Not observable in structural context.
Supplementary Signals
How this content communicates, beyond directional lean.
Learn more How well-sourced and evidence-based is this content?
0.72 medium claims
Sources 0.7 Evidence 0.7 Uncertainty 0.8 Purpose 0.8
No manipulative rhetoric detected
0 techniques detected
Emotional character: positive/negative, intensity, authority
measured
Valence +0.3 Arousal 0.4 Dominance 0.6
Does the content identify its author and disclose interests?
0.40
✗ Author
More signals: context, framing & audience Does this content offer solutions or only describe problems?
0.64 mixed
Whose perspectives are represented in this content?
0.50 2 perspectives
Speaks: institution
About: individuals community
Is this content looking backward, at the present, or forward?
mixed short term
What geographic area does this content cover?
unspecified How accessible is this content to a general audience?
technical high jargon domain specific
Longitudinal
1132 HN snapshots · 12 evals
Audit Trail
32 entries all eval pipeline all models llama-3.3-70b-wai llama-4-scout-wai deepseek-v3.2 claude-haiku-4-5-20251001
newest first
2026-02-28 14:22 eval_success Lite evaluated: Neutral (0.00) - - 2026-02-28 14:22
eval
Evaluated by llama-3.3-70b-wai : 0.00 (Neutral) reasoning Technical content no rights stance
2026-02-28 14:22 model_divergence Cross-model spread 0.56 exceeds threshold (4 models) - - 2026-02-26 23:19 eval_success Light evaluated: Neutral (0.00) - - 2026-02-26 23:19
eval
Evaluated by llama-4-scout-wai : 0.00 (Neutral) 2026-02-26 20:26 dlq Dead-lettered after 1 attempts: Learnings from 4 months of Image-Video VAE experiments - - 2026-02-26 20:24 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 20:23 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 20:22 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 17:47 dlq Dead-lettered after 1 attempts: Learnings from 4 months of Image-Video VAE experiments - - 2026-02-26 17:45 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 17:44 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 17:43 rate_limit OpenRouter rate limited (429) model=llama-3.3-70b - - 2026-02-26 14:39 eval_success Evaluated: Moderate positive (0.56) - - 2026-02-26 14:39
eval
Evaluated by deepseek-v3.2 : +0.56 (Moderate positive) 8,564 tokens 2026-02-26 09:19 dlq Dead-lettered after 1 attempts: Learnings from 4 months of Image-Video VAE experiments - - 2026-02-26 09:19 dlq Dead-lettered after 1 attempts: Learnings from 4 months of Image-Video VAE experiments - - 2026-02-26 09:18 rate_limit OpenRouter rate limited (429) model=mistral-small-3.1 - - 2026-02-26 09:17 rate_limit OpenRouter rate limited (429) model=hermes-3-405b - - 2026-02-26 09:17 rate_limit OpenRouter rate limited (429) model=mistral-small-3.1 - - 2026-02-26 09:16 rate_limit OpenRouter rate limited (429) model=hermes-3-405b - - 2026-02-26 09:15 rate_limit OpenRouter rate limited (429) model=hermes-3-405b - - 2026-02-26 09:15 rate_limit OpenRouter rate limited (429) model=mistral-small-3.1 - - 2026-02-26 00:57
eval
Evaluated by claude-haiku-4-5-20251001 : +0.39 (Neutral) 9,787 tokens -0.20 2026-02-26 00:33
eval
Evaluated by claude-haiku-4-5-20251001 : +0.58 (Moderate positive) 9,747 tokens +0.29 2026-02-26 00:05
eval
Evaluated by claude-haiku-4-5-20251001 : +0.29 (Mild positive) 10,070 tokens -0.15 2026-02-25 23:30
eval
Evaluated by claude-haiku-4-5-20251001 : +0.44 (Moderate positive) 9,594 tokens +0.01 2026-02-25 23:19
eval
Evaluated by claude-haiku-4-5-20251001 : +0.43 (Moderate positive) 9,995 tokens +0.01 2026-02-25 22:58
eval
Evaluated by claude-haiku-4-5-20251001 : +0.42 (Moderate positive) 9,736 tokens -0.03 2026-02-25 22:35
eval
Evaluated by claude-haiku-4-5-20251001 : +0.45 (Moderate positive) 6,909 tokens -0.04 2026-02-25 22:12
eval
Evaluated by claude-haiku-4-5-20251001 : +0.49 (Moderate positive) 7,020 tokens -0.08 2026-02-25 21:59
eval
Evaluated by claude-haiku-4-5-20251001 : +0.57 (Moderate positive) 6,661 tokens